4e:T1990,I've had a bug in Plinky since launch that I hadn't been able to fix until a few weeks ago. Every so often a paying user would see a banner asking them to upgrade to Plinky Pro, even though they'd already paid. Beyond it being a subpar user experience, I felt particularly bad because I was annoying someone who'd actively chosen to support my work. I tried quite a few fixes, including some approaches that I'm not particularly proud of — but none of them did the trick. Rather than trying another fix, I decided to see what would happen if I let Codex take a shot at the problem. As Codex was working I made a connection I'd never made before: whenever I hand my coding agent a few logs it gets dramatically better at debugging. LLMs are great at spotting patterns in large datasets, so I began to wonder — if 10 logs helped Codex figure out my bug, what would happen if I added 100, 1,000, or even 10,000? Agents Can't See What You See Anyone who's used a coding agent has had this very annoying experience: > You: Claude, please fix this [elaborately described] bug. > > Claude after churning for 10 minutes: I've fixed your bug! > > You: I just tried your fix and it doesn't quite work, and now this other thing is broken. > > Claude: You're absolutely right! Here's why it doesn't work, and here's a sonnet that describes why you're the most brilliant person I've ever met. Why does this happen over and over again? Because agents can't see what we see. They have no mental model of your app, no sense of coherence, and no inherent understanding of why they're doing this work in the first place. You know why you're solving a problem, but an LLM only knows what it was trained on and what lives in its context window. We hand LLMs tools to fill in the gaps — transforming models into agents. One such tool is automated testing, and it's a good tool, but it can only take you so far. When your agent writes tests for code that it doesn't innately understand it has no reliable way to know whether those tests are valuable. But agents aren't dumb. We know that they are remarkably good at observing patterns and anomalies in large batches of data. That's what our logs will be for — providing the agent a record of its work so it can pattern match against reality, rather than guessing what happened. What Are Logs For? The first program any developer writes is Hello World. The syntax changes from language to language, but it's always a variation on print("Hello World"). That line is a log at its simplest. We write a string, it shows up in the console, and we have confirmation that our code ran and did what we expected. The more an app grows, the less it looks like print("Hello World"). Nobody sits in a terminal watching text scroll by for fun, so what are logs actually for? Debugging. When a developer is trying to understand a problem they'll litter the code with temporary print statements. Adding print("Got here") helps them understand the flow of their program, and print("Synced links count:", links.count) will make sure the data in their app matches their expectations. When you read these logs back in order, you can see exactly where your mental model diverges from how the program actually ran. That divergence is a clue, helping you track down the root cause of the bug. Letting Logs Do Your Work With this understanding of logs, let's come back to Plinky and my subscription banner bug. Before Codex began working on a fix, I asked it to log out every action and state change related to subscriptions, then run through the app a few times to generate lots of logs. We then handed those logs back to Codex, and asked it to find the root cause. And it worked! Codex pored over an enormous pile of its own logs and found what I couldn't. The fix was creating a state machine to represent every subscription state change — including a few potential states I'd never noticed in my year of debugging this problem by hand. This was a better solution than anything I'd come up with, and the moment I realized I was onto something bigger. I began building an open source library to codify the pattern, and dogfooded it by logging every action and state change in Plinky. For the last month I've been handing Codex my bug reports, telling it to walk the code paths the user hit based on those logs, and having Codex fix the bug while I'm off doing something else. This works just as well on a greenfield project as it does on a year-old bug. I recently built myself a Spotify client in an afternoon, almost completely autonomously. I bootstrapped the project with a set of milestones and used Codex's /goal to keep Codex running until it finished all of them. Codex churned for three hours until I could log into Spotify, play my music, and do everything else you'd expect from a Spotify client. The reason it could do all of that unattended is that Codex wasn't just writing code anymore. The logs we added helped Codex determine what worked and what didn't in its own runs, even allowing me to watch Codex catch and fix its own bugs in realtime. From Bug To Library: Introducing Broadcast This wasn't a one-off result, and I've found the outcomes are so universally positive that this technique is something I now adopt in every project. That's why I codified this practice into an open source Swift[^1] library called Broadcast that integrates structured logging for humans and agents alike into any app. What started as a few logs in Plinky became thousands. And when I confirmed this strategy works, it became a library, a skill, and copious documentation that integrates Broadcast into any app — giving your agent the debugging superpowers it deserves. It still takes real thinking to build good software, but we're designing the strategies and tactics that let agents go further and faster than we could on our own. Last year I was staring at logs and my debugger trying to figure out why a bug was happening, while today I hand those logs to an agent to let it find the root cause. For Plinky, that means more features, built better, with fewer bugs. I spent a year unable to fix the race condition that caused some paying users to see upgrade banners, but Codex fixed it in an afternoon. It took weeks of work to lead to that afternoon though, because I had to give Codex the power to see what I couldn't make out with my own eyes. [^1]: I'll be porting this to TypeScript and any other language I use, but I wanted to get this right with good documentation before kicking that off.33:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mp7722u7cn2o",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/6/26/your-agent-deserves-logs/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering","Open Source","Swift"],"$type":"site.standard.document","title":"Your Agent Deserves Logs","description":"How structured logs helped Codex fix a year-old bug, and why logs are the key to unlocking autonomous workflows.","publishedAt":"2026-06-26T12:00:00.000Z","textContent":"$4e","canonicalUrl":"https://build.ms/2026/6/26/your-agent-deserves-logs/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mp7722u7cn2o","sort_date":"2026-06-26T12:00:00+00:00"}}] 4f:T122c,> A lot of people seem convinced that the point of AI coding is to write low-quality code as fast as possible. Spew out barely-passable slop, open massive PRs, and merge them unvetted. Ship it! > > But the thing is, LLMs are very flexible. And you can use them just as effectively to write high-quality code more slowly. > > This statement seems completely obvious to me at this point, and I almost didn't want to write this post for that reason. But there seem to be enough people convinced that LLMs are only good as slop cannons that it's worth making the opposite case. — Using AI to write better code more slowly --- Much like Nolan, I'd like to make the case against LLMs as slop cannons. Just because there are developers out there shipping 10,000 lines of code a day doesn't mean you have to be one of them. LLMs are a very flexible tool that you can mold how you see fit. If you prefer writing more high-quality code to shipping 10,000 lines of code a day, then you already have the tools to do so! If I had to describe my AI code writing process in one sentence it would be "vacillating between going really fast and slowing the heck down for a moment." I love using AI to build prototypes that validate my ideas at 10x the speed, but I can also build solid infrastructure at 1.5x the speed. I operate at every speed in between too. Some features are twice as fast, some fixes are five times faster. The trend line though is positive, which makes it a big boon to productivity. !Tiny gains graph showing 1% daily improvement compounding to 37.78x over a year People have trouble internalizing compounding growth, so I often reach for this chart to make it concrete. A 1% improvement every day adds up to nearly 3,800% over a year — which means you don't have to make every change a 10x leap to end up with 10x returns. Instead you can be excited about writing better code you can build upon, rather than more code that will eventually slow you down. I consider this a more balanced approach to AI coding. Sometimes it feels like Codex is omnipotent, but right now it's more useful to treat AI coding assistants as a potent pair programmer. A really smart developer who proposes great ideas, accepts my pushback, and fixes issues much faster than I would. I suspect we'll reach a point where it makes less sense for me to be this involved, but even if Codex never becomes more than a wildly useful coworker, we're all in for better software development experiences. Stop Focusing On Speed Speed is the main benefit everyone gets excited about with AI coding, but it's far from the only one: - Prototyping at lightning speed lets me validate or invalidate ideas much faster. People tend to think this means building features more quickly, but I also prototype API changes, speculative bug fixes, and even internal tooling I'd never otherwise justify building. I build a lot of internal tools these days that never would have existed, and the results compound. - If I explain myself well enough, I don't have to be stuck at a computer. I can take a walk and think more while Codex is writing code for me. Making forward progress used to mean sitting in my IDE in a state of flow, but with Codex Mobile I can check in on how things are going from anywhere. - My whole style of working has changed. I "code" in bursts by prompting, and while AI is writing code I get to focus on product design, support, and growth. These are all important parts of making a successful product, and I could never give them enough attention back when coding demanded my full focus. - I can contribute to products I normally wouldn't have the time or expertise to build from scratch. This weekend I submitted a pull request to Bluesky that adds the option to double tap your profile icon to switch accounts, much like Instagram or Threads. I've never written a line of React Native, but I'm a good enough developer to know what good code looks like. That qualifies me to make Bluesky a little better by guiding an agent — and to even improve on Codex's initial implementation to make sure the code is good. Quality > Quantity There's a lot more I love about AI coding, but that's a post of its own. The most important thing to remember is this: you can use AI to write 10x the code, or, if you care about code the way I do, you can use it to write better code. You can solve hard problems, improve your app's performance, or debug tricky race conditions. Being a 10x developer doesn't mean following the 10 latest trends on X. Being a 10x developer means leveraging the tools around you to become the best version of yourself, however you see fit.34:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaowu5z2d",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/5/26/being-a-1-5-10x-developer/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering","Productivity"],"$type":"site.standard.document","title":"Being A 1.5-10x Developer","description":"Good code begets good results, so let's make a case for using coding agents to write better code, rather than just more code.","publishedAt":"2026-05-26T12:00:00.000Z","textContent":"$4f","canonicalUrl":"https://build.ms/2026/5/26/being-a-1-5-10x-developer/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaowu5z2d","sort_date":"2026-05-26T12:00:00+00:00"}}] 50:T2226,Es ist so weit: Das Internet atmet tief ein. WordPress 7.0 (Codename „Armstrong“, benannt nach der Jazz-Legende Louis Armstrong) wurde offiziell veröffentlicht. Wenn im Netz eine neue Hauptversion der Software erscheint, auf der fast die Hälfte aller Websites weltweit läuft, geraten viele Hoster in Hektik. Es wird mit Features um sich geworfen, Hype-Feuerwerke werden gezündet – und ein seltsames „WordPress-Dooming“ bricht aus – die unbegründete Panik, dass ab morgen alles anders, komplizierter oder kaputt ist. Hier bei goneo in Minden trinken wir erst einmal einen Kaffee. Warum? Weil wir an „Boring Technology“ glauben. Die beste Technologie ist die, über die du im Alltag überhaupt nicht nachdenken musst. Ein Server muss laufen wie ein mechanisches Schweizer Uhrwerk – leise, präzise, unaufgeregt. Wir verändern nicht ständig deine gewohnte Umgebung, wir sichern sie einfach ab. Schauen wir uns also ganz pragmatisch, ehrlich und ohne Fachchinesisch an, was WordPress 7.0 wirklich für deine Website bedeutet, was sich im Alltag ändert – und warum du bei goneo absolut sicher bist. 1. Das Fundament zieht an: PHP 7.4 ist die neue Untergrenze WordPress 7.0 hat die alte PHP-Mindestanforderung angehoben. Wer noch auf ganz alten PHP-Versionen funkt, steht ab heute vor einer verschlossenen Tür. Die goneo-Lösung: Während andere Hoster jetzt automatisierte Warn-Mails verschicken, die klingen, als würde morgen die Welt untergehen, sagen wir: Entspann dich. Bei goneo laufen unsere Systeme ohnehin längst auf den modernen, pfeilschnellen PHP 8.x-Versionen. Falls deine Website doch noch einen alten Schubser braucht, führt dich unser goneo-Kundencenter elegant und einfach per Klick zum Update. Technik menschlich erklärt, Schritt für Schritt. 2. Im Fokus: Die neuen „DataViews“ (Dein aufgeräumter Arbeitsplatz) Die größte und spürbarste Änderung im Alltag betrifft die Art und Weise, wie du deine Inhalte verwaltest. Die Entwickler haben die alten, starren Admin-Tabellen im Hintergrund durch ein modernes System namens DataViews ersetzt. Das ist ein echtes Frontend-Evolution-Overlay, das dir die tägliche Arbeit massiv erleichtert, anstatt dich zu überfordern. Schluss mit starren Tabellen: Mit den neuen DataViews filterst du deine Beiträge und z.B. SEO-Werte im WordPress-Backend flüssig und in Echtzeit – ganz ohne lästiges Neuladen der Seite. Stell dir DataViews wie eine perfekt sortierte, digitale Werkzeugwand vor. Du findest diese neuen Ansichten direkt im WordPress-Adminbereich, vor allem im Site-Editor unter Seiten, Beiträgen oder Vorlagen. Drei Dinge machen das Feature extrem sinnvoll für deine Routine: Visueller Überblick (Das Raster-Layout): Wenn du deine alten Blogbeiträge oder Seiten überarbeiten willst, musstest du früher jeden Titel einzeln anklicken, um das Beitragsbild zu sehen. Jetzt schaltest du einfach auf das Raster-Layout (Grid) um. Du siehst deine Beiträge wie in einer Galerie und erkennst sofort, wo ein Bild fehlt oder das Format aus der Reihe tanzt. Filter in Echtzeit: Du suchst alle Entwürfe eines bestimmten Autors aus einer bestimmten Kategorie? Ein Klick auf „Filter“, Kriterien auswählen – und die Liste passt sich sofort an, ohne dass die gesamte Seite träge neu laden muss. Massenbearbeitung ohne Frust: Zehn Seiten gleichzeitig aktualisieren oder kategorisieren geht jetzt flüssig von der Hand. Die Aktionsleiste arbeitet im Hintergrund so geschmeidig, dass du Dutzende Beiträge in Sekunden organisierst. Du kannst dir sogar dein ganz persönliches Dashboard zusammenstellen, unbemerktes Chaos aufräumen und diese Ansicht für morgen speichern. 3. KI zieht in den Kern ein – aber wir bleiben auf dem Teppich Großer Hype im Vorfeld: WordPress 7.0 bringt einen integrierten „WP AI Client“ mit, eine standardisierte Schnittstelle für KI-Plugins (wie für Google, OpenAI oder Anthropic). Unsere Haltung dazu: Wir hosten deine Daten zu 100 % mit Ökostrom in unserem Rechenzentrum in Frankfurt. Das ist deine digitale Souveränität. Wenn du KI-Features nutzen möchtest, kannst du das tun – die Rechenlast und die API-Kosten liegen bei den externen Anbietern, deine goneo-Server-Ressourcen bleiben davon völlig unberührt. Keine Kostenfallen, kein unberechenbares Cloud-Billing. Bei uns gilt: Feste Ressourcen, transparente Fixpreise. Der neue WP AI Client in WordPress 7.0: Hier lässt sich modular einstellen, welche KI-Assistenzfunktionen – wie die automatische Textzusammenfassung oder Alt-Text-Generierung 4. Was nicht drin ist: Echtzeit-Kollaboration verschoben Eigentlich sollte mit Version 7.0 das gemeinsame, gleichzeitige Bearbeiten von Texten kommen, wie man es von Google Docs kennt. Zwölf Tage vor dem Release hat das WordPress-Kernteam die Reißleine gezogen – zu viele Bugs, zu hoher Server-Load (goneo easyPage, unser Website Builder, kann das übrigens schon lange: Man kann andere User per Mail zum Echtzeit-Mitarbeiten einladen). Unser ehrliches Fazit als Hoster: Danke für diese Vernunft! Unreife Features, die Server wahllos in die Knie zwingen, nur um einen Hype zu bedienen, passen nicht zu unserer Philosophie. Es ist besser, ein Feature zu verschieben, als die Stabilität deiner Live-Website zu riskieren. Was musst du heute tun? Wenn du deine Website bei goneo hast: Erst einmal nichts. Wir überwachen die Infrastruktur rund um die Uhr. WordPress 7.0 wird sich in den kommenden Tagen organisch und sicher verbreiten. Wenn du automatische Updates aktiviert hast, wird deine Seite das Update erhalten. Weil wir die technischen Hürden im Hintergrund flach halten, läuft deine gewohnte Umgebung einfach stabil weiter. Lass dich vom WordPress-Dooming im Netz nicht ablenken. Genieße den Tag, kümmere dich um dein Handwerk und deine Inhalte – das unaufgeregte goneo-Systematik im Hintergrund läuft verlässlich für dich. Kurzantworten für den Kaffee danach: Das WordPress 7.0 Q&A Kostet mich der neue KI-Client bei goneo extra? Nein. Der „WP AI Client“ ist ein reines Software-Werkzeug im WordPress-Kern. Wenn du ihn nutzt (indem du ihn mit Anbietern wie OpenAI, Claude, Jetpack oder Google verknüpfst), rechnet der jeweilige KI-Anbieter seine API-Gebühren direkt mit dir ab. Auf deine goneo-Hosting-Gebühren hat das absolut keinen Einfluss. Bei uns bleibt es bei deinem fairen, transparenten Fixpreis. Was passiert, wenn meine Website noch auf einer alten PHP-Version läuft? Wenn deine Website noch auf einer Version älter als PHP 7.4 läuft (was bei goneo nicht mehr sein kann), verweigert WordPress 7.0 das Update, um deine Seite vor Fehlern zu schützen. Bei goneo bist du aber auf der sicheren Seite: Unsere Standard-Infrastruktur läuft längst auf den schnellen Versionen PHP 8.x. Falls deine Seite einen Schubs braucht, kannst du deine PHP-Version im goneo-Kundencenter mit nur einem Klick ganz einfach und sicher selbst anheben. Muss ich die neuen DataViews jetzt zwingend nutzen? Keine Sorge, du musst nichts Neues lernen, wenn du nicht willst. Die DataViews ersetzen zwar nach und nach die alten, starren Tabellen im Hintergrund, aber das System steuert sich komplett intuitiv. Es ist wie ein aufgeräumter Schreibtisch: Du siehst deine Beiträge und Seiten sofort, kannst sie aber viel schneller filtern, sortieren oder als Raster anzeigen lassen. Es nimmt dir keine Funktionen weg, es spart dir nur Klicks. Wo ist die Echtzeit-Kollaboration abgeblieben? Das gemeinsame Schreiben wie in Google Docs oder auchin unserem easyPage wurde kurz vor dem Release aus Version 7.0 gestrichen, da es zu viele Server-Ressourcen gefressen hätte und noch zu fehleranfällig war. Das WordPress-Team entwickelt das Feature im Hintergrund weiter. Wir als Hoster begrüßen diese Entscheidung: Stabilität geht immer vor Hype. Deine Website blei Kann durch das Update meine Website abstürzen? Da WordPress 7.0 tiefgreifende Änderungen an den Admin-Oberflächen (DataViews) vornimmt, könnten extrem veraltete Plugins, die sich in diese Bereiche einklinken, kurz ruckeln. Wir empfehlen: Aktualisiere vor dem großen Sprung all deine Plugins und Themes. Und falls doch mal etwas hakt? Keine Panik. Unser menschlicher Support sitzt in Minden und hilft dir direkt weiter – ganz ohne sterile KI-Warteschleife. Welche Version von WordPress bekomme ich installiert, wenn ich das Paket WordPress-Power nutze? Es wird grundsätzlich die aktuellste Version automatisch installiert, nun als WordPress 7.0. Ab diesem Zeitpunkt entscheidest du, wie die Updates einer bestehenden Installation vornimmst. Entweder du stellst alles auf „Updates automatisch installieren“ oder du klickst die entsprechenden Menüpunkte für den WordPress-Kern (Core), die Themes und Plugins.35:["$","$L30","at://did:plc:zzy2bg27vikelsirgrnz4tlw/site.standard.document/3mmwf2tas5vaz",{"publication":{"href":"https://social.goneo.de","pubRecord":{"url":"https://social.goneo.de","icon":{"ref":{"$link":"bafkreic5hpvogoev3zp4yz5tqemfm2sycsg6a4ic3cebj5zxljz4kuhg6u"},"size":65126,"$type":"blob","mimeType":"image/png"},"name":"social.goneo.de","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":88,"g":238,"r":255,"$type":"site.standard.theme.color#rgb"},"background":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":17,"g":17,"r":17,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":0,"g":0,"r":0,"$type":"site.standard.theme.color#rgb"}},"description":"goneo Social Content Hub","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showFirstLast":false,"showInDiscover":true,"showRecommends":true,"prevNextDirection":"rtl"},"theme":"$undefined"},"uri":"at://did:plc:zzy2bg27vikelsirgrnz4tlw/site.standard.publication/3mmwex72ga2bm"},"ownerDid":"@goneo.bsky.social","contributors":[{"did":"did:plc:zzy2bg27vikelsirgrnz4tlw","handle":"goneo.bsky.social","displayName":"goneo"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/05/21/wordpress-7-0-mit-dataviews/","site":"at://did:plc:zzy2bg27vikelsirgrnz4tlw/site.standard.publication/3mmwex72ga2bm","tags":["Technologien bei goneo","top","Wordpress"],"$type":"site.standard.document","title":"WordPress 7.0 „Armstrong“ ist da: Warum das beste Update das ist, von dem du kaum etwas merkst.","updatedAt":"2026-05-21T10:55:11.000Z","coverImage":{"ref":{"$link":"bafkreifxtxkg6cq2ffib3ycbyunv2yox2mgxgswfcnuxrk5xshowimtwjm"},"size":594855,"$type":"blob","mimeType":"image/jpeg"},"bskyPostRef":{"cid":"bafyreiea7d5st3gsrq4vkirbbpjtvkrc2shvnlegvqdadw6wc4zubh3bq4","uri":"at://did:plc:zzy2bg27vikelsirgrnz4tlw/app.bsky.feed.post/3mmwf2taltnaz"},"description":"Das Internet diskutiert über WordPress 7.0 und den neuen Update-Stress. Wir bei goneo trinken erst einmal einen Kaffee. Warum das neue Feature „DataViews“ deinen Arbeitsplatz im Backend revolutioniert, warum der KI-Hype uns völlig kaltlässt und weshalb deine Website bei uns so sicher läuft wie ein mechanisches Uhrwerk.","publishedAt":"2026-05-21T10:52:55.000Z","textContent":"$50","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:zzy2bg27vikelsirgrnz4tlw/site.standard.document/3mmwf2tas5vaz","sort_date":"2026-05-21T10:52:55+00:00"}}] 51:T1929,I was refactoring Plinky's deeplink routing layer, and the process led to an unbelievable computing experience. Well it would be more accurate to say that Codex was refactoring Plinky's deeplink routing layer while I slept, and then I had an unbelievable computing experience. !The Plinky Small Wins doc In my recent talk at Deep Dish Swift I shared that I keep a doc called Plinky Small Wins. This is where I keep small tasks that I want Codex to do overnight, and I wake up to a manageable pull request that I can review. I don't want the work to be too big because then I'll have a lot of code to review — and I'm still the blocker for shipping code to production. I consider this constraint a feature, because it allows me to maintain quality control while benefitting from an agent that works while I sleep. This strategy works remarkably well because AI produces great results when given a well-defined narrow task to execute on. There's always a test to write, some code to clean up, or something to improve, and there's minimal risk since I'm reviewing everything before it ships. !The Low Hanging Fruit automation running in Codex Every morning at 7am Codex runs this Low Hanging Fruit automation and reads the Plinky Small Wins doc over MCP so when I wake up I have some code waiting for me. This morning Codex refactored Plinky's deeplink router and delivered a solution that was about 80% of what I wanted. The result was pretty much what I would have done, with a few changes that were actually better than anything I'd considered. There was one assumption that drove an architectural change I disagreed with, but it was easy enough to ask Codex to correct that. I'm still the human in the loop, guiding the computer to the code I want. As a result I still need to review the generated code, tests, and validate the work my agent did. Putting The Computer To Use The code looked very good (and there was one idea that I really loved). The tests Codex wrote looked reasonable as well. But I was dreading the painfully boring task of manually clicking through a bunch of deeplinks to make sure they still opened to the right screens in Plinky. So I thought to myself, what if I could make the computer do that instead? Lately I've been low-key obsessed with Codex's new Computer Use plugin. Codex is unbelievably good at using a computer to do practically any task you throw at it. It's also a terribly inefficient way to use a computer, except on one axis: the computer is working on your specified task while you're doing something else. I don't recommend Computer Use for everything, but it's perfect for mindless, repetitive work. And so I wrote this prompt to send Codex off to use the iOS simulator to make sure none of my deeplinks had broken: > While I'm reviewing the code, can you using @Flowdeck and @Computer to run a simulator and test to make sure we haven't broken any routes? I want you to open Safari with deeplink URLs. You don't have to test every route, but you should test at least 5, selecting the ones that are most likely to be broken through our changes. Upon opening Plinky via deeplink I'd like for you to wait 3 seconds and to take a screenshot. This will allow us to verify that the screen we expected to open did in fact open, or that the action we're firing off worked. Beyond My Intuition Computer (as I'll refer to Computer Use going forward) did in fact do what I asked it to do, but not quite how I expected. As you can imagine Computer is not human, so it couldn't navigate the iOS simulator with as much dexterity as I can. It ran into a problem typing out custom URL schemes thanks to some autocorrections in my iOS dictionary. Computer then pivoted and tried copying and pasting the deeplink string instead — but that backfired by executing a Google search for the URL. There were more hacky paths to take like tapping into iOS's accessibility features, but we never got there. !Multiple iOS simulators open during the deeplink testing process Instead Computer did something brilliant I'd never thought of: It created a small HTML page in /tmp with six deeplinks into Plinky that it could click to test our deeplinks. Rather than wrestling with the iOS simulator's constraints, it sidestepped the problem entirely. After that, testing our functionality was incredibly easy, as you can see by the results. Thinking Differently This wasn't just an example of automation making my life easier, it was also a moment where AI taught me how to be better at my job. In retrospect I should have a fixture like this in my app that I can reference, rather than the manual approach I've always taken for testing deeplinks. Over the last few weeks Codex has built custom CLIs for Plinky I've always wanted, suggested better approaches to problems I'm solving, and built a feature for Boutique I've wanted for over three years. Today's experience reminded me of an interesting note I found in the GPT-5.5 usage guide: > Improved instruction following: GPT-5.5 interprets prompts in a literal and thorough manner, enabling specific, descriptive instructions when the product requires them. Define success criteria and stopping rules, especially for long-running, tool-heavy, or evidence-gathering workflows. See: Write outcome-first prompts and Keep the right specificity. This may seem like a subtle change, but the implications are radical. Because GPT-5.5 takes your constraints literally, the clearer you define the boundaries you need respected and your success criteria, the better the results will be. Many people have begun treating their job as telling AI what to do, but you'll get much better success working with AI if you focus your time on assembling context. Rather than providing step by step instructions, your job is to define the solution space your agent should work in. When you do that your agent will find paths you hadn't considered, the same way I hadn't thought to build out a deeplinks page as testing infrastructure. And I bet you'll learn things about your own work too. The term "prompt engineering" felt silly when it was coined because anyone can write words, so it disappeared from the lexicon as quickly as it appeared. Anyone can write words, but providing the right context is a real skill that not everyone excels at — and it's the skill that matters most today. The best thing you can do right now is to get better at defining that space.36:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqap4tka2i",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/5/11/the-teaching-computer/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering","Plinky","Productivity"],"$type":"site.standard.document","title":"The Teaching Computer","description":"Many people are using AI to solve their problems when they could be leveraging AI to become better versions of who they want to be.","publishedAt":"2026-05-11T12:00:00.000Z","textContent":"$51","canonicalUrl":"https://build.ms/2026/5/11/the-teaching-computer/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqap4tka2i","sort_date":"2026-05-11T12:00:00+00:00"}}] 52:T1bce,Much of the tech world is gushing about the accidental leak of Claude Code's source code yesterday, but for different reasons than I find it interesting. I began jotting down my thoughts and came up with five distinct observations that had little to do with the leak itself, and more about what it tells us. 1. The Code Is Garbage !"Tired: Omg the Claude Code leak is a bunch of vibe coded garbage Wired: Vibe coded garbage can get you to $2.5 billion annualized recurring revenue in under a year if the product market fit is there"

Tired: Omg the Claude Code leak is a bunch of vibe coded garbage Claude Code is a beloved product, to the point where developers, designers, product managers, marketers, and even CEOs are obsessed with it! And yet the code that powers Claude Code is kind of garbage. So of course the first thing people did was point and laugh. But step back for a second and think, what does that tell us about the actual value of code? I argued in AI Agents Are Starting To Eat SaaS (Really) that the barrier to entry for creating a product is going down. That seems like a statement about toy apps like todo lists and habit trackers — but it applies to all software. The success of Claude Code and Cursor at the higher end of the market shows that even the people pickiest about their software (developers) will use your software regardless of how good the code is. Many software developers have argued that working like a pack of hyenas and shipping hundreds of commits a day without reading your code is an unsustainable way to build valuable software, but this leak suggests that maybe this isn't true — bad code can build well-regarded products. 2. It's Not About The Code It should serve as a warning to developers that the code doesn't seem to matter, even in a product built for developers. This interview with Boris Cherny (the creator of Claude Code) was eye-opening for me. He describes how they build software at Anthropic and explains why the code matters — just not in the way developers typically assume. What matters is what the code does, not how it does it at the character-by-character level. Anthropic isn't only building better systems to write better code, they're building better observability systems to monitor the effects of code changes. Imagine you've built a feature and now it's time to QA it. You notice that an email textfield doesn't respond well to the @ character, so you go back to the code, read it, and with enough debugging you figure out a fix. But that doesn't scale as well as a system that yells at you to say "users can't log in right now", and then goes back to automatically change or revert the code that broke your auth flow. If you can build a good self-healing system and are willing to take on a little risk of things breaking as you go, you can move a whole lot faster — not just a bit. 3. It's About Product Market Fit As always, product market fit is the only thing users care about. If the product works, very few people care how it works under the hood. Heck, most people don't even have an inkling of what's actually happening behind the scenes. There's always a chance that Claude goes to shit (or just goes down every day because Anthropic's servers are under-provisioned due to poor demand prediction). If that happens, OpenAI can jump in with their equally good (if not better) model and leverage the ridiculous amount of servers they have to serve the latent demand. Or maybe Google will eventually figure out how to ship a good coding product. There's plenty of opportunity here, and ultimately we're supply-constrained in meeting consumer demand. 4. Copyright Is Still A Touchy Subject The whole copyright situation is very funny to me, and feels a bit like Anthropic is getting a taste of their own medicine. But I think there's more to it than just what comes around goes around. The first thing Anthropic did when their code leaked was send a bunch of DMCA notices on Github to have the repos taken down. True to their committment to vibing, Anthropic ended up sending DMCA notices to forks of their own claude-code repo that hosts their skills, tutorials, and example code. But then the clean room implementations started showing up. People had taken Anthropic's source code and rewritten Claude Code from scratch in other languages like Python and Rust. The whole AI industry — Anthropic included — has been arguing that using AI to rewrite something is not derivative work and doesn't violate copyright, because that is how they themselves train their models. Now this part really does feel like Anthropic's getting a taste of their own medicine. But my higher-level reading is that this further entrenches the idea that code should be free, just with a more libertarian bent than the Free Software Foundation expected. 5. This All Doesn't Matter All of this is interesting, but I think Claude Code's source code being leaked won't matter as much as people seem to think it will. The real value in the AI ecosystem isn't the model or the harness — it's the integration of both working seamlessly together. Anthropic could open source Claude Code tomorrow and it wouldn't change a thing, because what people are paying for is the great results, not the underlying code. Codex has been open source since launch, and Gemini is too. Neither has captured Claude Code's mindshare even though many people prefer Codex — because what Anthropic is selling is a complete service. Lately I've been using the pi coding agent a lot, and I love it. Pi is a coding agent with just four tools: read, write, edit, and bash. It works with every major model provider — including Claude — and it works brilliantly. The reason is that it's optimized for working through problems the way a developer would solve them — by writing code. This is a different approach to Claude's abundance of tools, which goes to show there's a diversity of ways to create an integrated experience across model and harness. So… Where Does That Leave Us? I've had to question the value of code a lot over the last couple of years, and this leak continues to reinforce the notion that I've vastly overestimated it my entire career. What matters is integration. Whether that's product market fit or how well a model and harness work together, users have always cared about having their problems solved — solved well, really. You can build something great by making it simple or complex, open or proprietary, but it has to work seamlessly. A clean codebase only matters if it delivers better results for users. This leak changes the perception of Claude Code more than it changes anything tangible, but perception is reality. And the reality is that the code was never what made Claude Code valuable in the first place — everything happening around the code matters more.37:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqap7lpi2z",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/4/1/the-claude-code-leak/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Engineering","Society"],"$type":"site.standard.document","title":"The Claude Code Leak","description":"What the accidental Claude Code source code leak tells us about the real value of code, product market fit, and why integration is what actually makes software great.","publishedAt":"2026-04-01T12:00:00.000Z","textContent":"$52","canonicalUrl":"https://build.ms/2026/4/1/the-claude-code-leak/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqap7lpi2z","sort_date":"2026-04-01T12:00:00+00:00"}}] 53:Tef4,If you want the 7 word version of this blog post it’s You are responsible for what you ship. And yet, I’m constantly amazed by how many software developers, product designers, and CEOs seem to forget this the first time they see Claude build something complex and impressive in almost no time — not considering the incorrect subtleties. We’re currently living through a paradoxical moment. AI can do remarkable things like write high quality code, college-level essays, or consultant-quality research — in minutes. These tasks previously would have taken hours, but that speed comes with a hidden cost for those that aren’t paying attention. Even the best code has bugs, college students don’t write perfect essays, and consultants make plenty of mistakes too. And yet when the output comes from Claude, people entirely skip over the rigor they used to apply. Expectations vs. Reality I’ve been working with and teaching AI for over 3 years. When I talk to people about AI as a tremendous accelerant, they’ll inevitably share an experience they had where a coworker submitted a bunch of AI slop that they didn’t write, check, or test. And that leads to the question: what do I do if my coworker generates some AI code (or writing or whatever) that isn’t very good? And I always respond that even in an era of AI the whole point of your work is that it still has to work! The fact that Claude can generate something in minutes doesn’t lower the bar for quality — even if agents are doing increasingly amazing things by the day. If you want to continue using agents then you need to focus on improving their outputs, not gleefully or begrudgingly accepting whatever they spit out and moving on. Before Claude (BC) If a coworker submitted bad code before Claude Code what would you have done? You would have pointed out the problems, asked them to fix it, and you wouldn’t have accepted it into your codebase until the code was ready to ship. Even if we can more quickly reach the bar for shipping, the bar hasn’t been lowered. If anything, it should be higher now that we can have AI assist with and validate our work. If your coworker continues submitting slop, you have to tell them it’s not acceptable. If they don’t understand why, then this person probably isn’t ready for an environment where agents write code. Agentic coding is a new way of working for everyone, so a good employer will invest in their training. A less generous company will simply do what they’ve always done — fire the underperforming employee. The Claude Era (CE) Having AI available isn’t an excuse to think less, it’s an opportunity to work smarter. When you don’t check the work AI does for you the only thing you’re doing is creating work for others to review. How would you feel if someone gave you more work to do because they were too lazy to do theirs? I expect the more common this negative feedback loop becomes the more high-functioning teams will simply give it a hard nope and tell someone to start over. The alternative is untenable and only reduces productivity, so eventually better judgment will prevail. For years, many in our industry mused that “code doesn’t really matter — it’s all about the user” and stated that “software is a team sport”. If you believed that then and believe it now, nothing has changed about what it takes to ship software. As Simon Willison so succinctly states: “Your job is to deliver code you have proven to work.” The Harvard Business Review correctly identifies that workslop is on the rise, so I’ll leave you with three rules for sloppy work. 1. Don’t be the person who supports workslop. 2. Don’t be the person who accepts workslop. 3. And most importantly: don’t be the person who creates workslop for others.38:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapbzlj2d",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/3/23/workslop/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Programming","Engineering"],"$type":"site.standard.document","title":"Workslop","description":"The bar for shipping quality work hasn‘t lowered just because AI can generate code in minutes — and if you‘re accepting AI output without reviewing it, you‘re part of the problem.","publishedAt":"2026-03-23T12:00:00.000Z","textContent":"$53","canonicalUrl":"https://build.ms/2026/3/23/workslop/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapbzlj2d","sort_date":"2026-03-23T12:00:00+00:00"}}] 54:Tc4b,I've just released Recap 2.0 — my open source library for creating beautiful What's New screens from markdown. The primary purpose for this update was interface improvements I need to include in Plinky for Mac (available very soon!), but the underlying story is more important. This is my first open source release where an agent wrote 100% of the code. I was deeply involved in prompting and guiding the work, which taught me an important lesson about working with agents. Building For Agents I write code with agents every day — probably 90-95% of my code is the result of an agentic workflow. When I started working on Recap that number was closer to 5 or 10%, so my working style has changed a lot more than the project itself. When I started updating Recap for 2.0, I quickly realized the codebase wasn't set up well for agents to work with. Updating Recap felt a bit like stepping into a time machine because the project had no structure in place to make agentic development easier. The first thing I did was create an llms.txt file that defines the project's structure and functionality. This small file makes a big difference, letting an agent be laser-focused rather than fumbling around brute forcing solutions. Then I built a Recap Integration Skill that handles the hard work of adding Recap to your project. Something I've learned as an open source maintainer is that people will find the most creative ways to misuse your library. I've always wanted a way to prevent questions from people who've misunderstood how my open source project works and gotten stuck trying to fix their mistakes. In the past I've solved this by writing lots of time-consuming documentation. The tradeoff is worthwhile for my projects like Boutique which have thousands of users, but there isn't as much payoff for smaller projects. Creating a skill solves this by teaching an agent how to write the code you want, and that's what makes the recap-integration skill the real highlight of Recap 2.0. Updating a project for agentic workflows isn't just for open source, it will help you in any codebase where you use tools like Claude Code, Codex, and Cursor. Your Turn If you want to get better at working with agents, start small. Some of the highest leverage work you can do is to make your codebase agent-friendly. Much like professional pianists still practice their scales, you can level up your fundamentals by practicing on a small project. It's more difficult to begin with a large established codebase, but the lessons you learn in a small project will transfer over to codebases of all sizes. The work of making your codebase agent-friendly starts with one small question: What would make your codebase easier for an agent to work with? You can figure this out by reading up on agents, taking a workshop, or even trial and error — but that question alone will point you towards the changes you need to make. If you want to see what that looks like in practice, check out Recap on Github. It's a concrete example of a small project restructured with agents in mind, and I'm always happy to answer any questions you may have about it.39:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapeior2d",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/3/11/small-steps-for-agent-friendly-codebases/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering","iOS","Swift","Open Source"],"$type":"site.standard.document","title":"Small Steps For Agent-Friendly Codebases","description":"Updating an open source library taught me a valuable lesson for making your codebase agent-friendly — small steps can have a big payoff.","publishedAt":"2026-03-11T12:00:00.000Z","textContent":"$54","canonicalUrl":"https://build.ms/2026/3/11/small-steps-for-agent-friendly-codebases/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapeior2d","sort_date":"2026-03-11T12:00:00+00:00"}}] 55:T1843,import communicationDemoHtml from './communication-demo.html?raw'; export const communicationDemoUrl = data:text/html;charset=utf-8,${encodeURIComponent(communicationDemoHtml)}; > I think Twitter is as guilty of hiring too fast as Google and most of these other companies, the problem of bloat is just way bigger at Google and has been for over a decade now. That's what I said in 2020 while working at Twitter, and I still believe it. Now as I watch companies like Square lay off 40% of their workforce, I'm certain now that this isn't solely about corporate greed or AI replacing workers — this is a sectoral shift that's been a decade in the making, and we're finally seeing it accelerate. People who have never worked at a large tech company believe that these companies are overstaffed. But workers at these companies will tell you the opposite — convinced their firm is understaffed because they're constantly fighting for resources. So how do we square this circle? This isn't directly a symptom of mismanagement, mismanagement is what happens when you underrate the hidden costs of communication. Size Matters These moves are rational because large companies aren't overstaffed or understaffed — they're overstaffed and under-focused. That's why everyone at tech companies with 50,000 employees are always complaining that despite 49,999 coworkers, everything is inefficient. The naive solution is to hire more, but the real answer is to reduce bandwidth. There's a saying that the job of a CEO is to say the same thing 10 times until everyone hears it. The larger your company, the further information has to diffuse, the more people have to coordinate — and the harder it becomes to actually get anything done. Go ahead and try it out yourself.
</div> Notice that when we go from 1 to 5 people we don't just have 4 more people, we have 10 new connections between people. Going from 5 to 10 increases connections from 10 to 45. Going from 10 to 500 brings us to 124,750 connections, and 50,000 people takes us all the way to 1,249,975,000 connections. Building a product is a team sport, and it gets exponentially harder to play together the more people you add. Communication Is Everything Imagine you're a CEO and need to make a fundamental shift in your product: - 1 person: You shift the product instantly. - 5 people: You get everyone in a room and decide in a few hours. - 10 people: You do all that and hold a company meeting where you explain why you're making this change. - 50 people: You coordinate with your executive team on how to cascade this change through the organization. - 500 people: You reshape the entire company around this change — layoffs, restructuring, and a redefined mission. - 50,000 people: You execute this across every organization, team, and product, and make your case to public markets. This is why the goal of a company is not to grow, but to avoid headcount growth while maximizing revenue and profits. Because the overhead for communication is so high, every employee you hire only produces marginal returns — maybe 10 or 20% more throughput once you account for coordination costs. AI It's hard to ignore the elephant in the room: AI. AI really can make people more efficient and allow them to do more with less. But as this HBR study shows, AI doesn't reduce work — it intensifies it. The bottleneck was never the amount of work a person could do. It was always the number of interconnected edges in a system — the communication overhead. AI makes individuals more productive, but it doesn't make large organizations better at coordinating. The people don't get better at communicating as fast as the models get better at doing the work. So you can do more, but you still can't align 50,000 people any faster than you could before. This is why the layoffs we're seeing aren't just about replacing workers with AI. They're about companies finally accepting what was always inevitable — you don't need 50,000 people to do the work of 30,000. You never did, but AI has made this reality impossible to ignore. Workers Now don't get me wrong, this essay isn't meant to be some sort of pro-capital anti-labor screed. My heart goes out to all of the people who were laid off from Square yesterday, from Amazon before that, and from Microsoft, Google, and all the other big tech companies that have done layoffs recently. This all makes me very squirmy as a person who sometimes considers reentering the workforce. This post isn't meant to be solace, but a cold hard look at our reality. For companies like Square, this is a reset to focus on what matters most — with fewer employees and more efficiency. I've been saying this for 6 years — long before ChatGPT — so I'm confident this isn't just an AI story. In response, I always hear an optimistic argument — these smart people will start companies of their own and outcompete the big players. But there's no evidence the market needs more companies selling the same products to fewer employed people. If 50 people get laid off from TikTok, there's no natural law that says the market will suddenly clamor for more innovation in short-form video. Some of these people will certainly create new companies and new jobs. But there's no guarantee those jobs will be recouped — and even if they are, it may take years. In the meantime, the people laid off today are looking for work in a market that's contracting, not expanding. And that’s why my takeaway is that big companies are going to get smaller and smaller, and smaller companies may follow suit. I wish I had tangible advice to share, but all I can say is brace for change across the industry, because it’s a-comin’.3a:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapj54r2d",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/2/27/the-hidden-cost-of-communication/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry"],"$type":"site.standard.document","title":"The Hidden Cost Of Communication (A Case Against Big Companies)","description":"Big companies are downsizing, not solely because of AI and corporate greed but because communication bottlenecks kill productivity and demoralize teams.","publishedAt":"2026-02-27T12:00:00.000Z","textContent":"$55","canonicalUrl":"https://build.ms/2026/2/27/the-hidden-cost-of-communication/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapj54r2d","sort_date":"2026-02-27T12:00:00+00:00"}}] 56:T2726,There's an adage that says using AI is like having an intern at your disposal. The metaphor usually carries a negative connotation — a lowly helper without much skill. But it's apt because of the style of work AI actually does. 1. You set an intern off to go do some work 2. They come back saying they've completed their assignment 3. You show them where there are gaps, and nudge them in the right direction 4. Repeat 5-10x until they get where you wanted them to go Last week OpenAI released an app for their coding agent Codex. [^1] At first glance, it looks like it just makes their Codex CLI more accessible — much like Claude Cowork is an interface to Claude Code for people who don't live in a terminal. But there's a premiere feature that instantly provided value: automations. Automations An automation in Codex is as simple as it sounds. You configure a task to run on a regular basis, and it goes off and does some work for you. This work can be anything — and it's safer than other approaches because it runs in whatever sandbox environment you've set up. Setting up an automation is easy. You give Codex a title, decide which projects it can access, and write a prompt to execute. Then you pick how often the automation will run — and it'll execute whenever your computer is on. (In the future, automations will run in Codex Cloud, so you won't need to keep your computer on.) !An automation configured in Codex Your Summer Intern Let's push the intern analogy a little further. On their first day, an intern knows nothing and your job is to teach them the ropes. They don't know anyone, don't know what they'll be doing, and don't even know how you take your coffee. You should expect little from a college kid who's never held down a full-time job before. By summer's end though, they'll have experience under their belt. They know their projects, understand how to work with product and design, and you can leave them alone for longer stretches of time. If you've been working closely with them, they may even anticipate your next move. This anticipation is what automations do so well. If you've defined a task well, you suddenly have an intern who understands what you need. Not only can they do it, but they're confident and now proactive about it. This is the agentic unlock that automations bring — what really separates an automation from using ChatGPT to solve a problem. Real World Automations My first automation was inspired by a tweet from Dominik Kundel of OpenAI's SDK team: <blockquote class="twitter-tweet" data-dnt="true"><p lang="en" dir="ltr">💪 Have Codex automatically become better<br><br>Add a daily task to read your past sessions in ~/.codex/sessions and update your <a href="https://t.co/DKdqY6VTPu">https://t.co/DKdqY6VTPu</a> files or update/create new skills for common tasks. <a href="https://t.co/8ME7LWu3ML">pic.twitter.com/8ME7LWu3ML</a></p>— dominik kundel (@dkundel) <a href="https://twitter.com/dkundel/status/2018455599391121851?ref_src=twsrc%5Etfw">February 2, 2026</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script> That got me thinking: what else could I make Codex do while I slept? I wanted something small enough to easily review, but important enough to actually derive value from. Basically, I want someone to help me with all the low-hanging fruit on my roadmap — so that's exactly the automation I set up. Low Hanging Fruit I built a simple automation that works like this: 1. Codex reads Plinky's internal roadmap, which I keep in a Craft doc. This document is accessible via MCP, so Codex can read through my backlog of features and issues. 2. Codex reads my codebase to understand how easy or hard each task will be. 3. Then it builds a small feature, improves code quality, or fixes up some technical debt. I turned those steps into a prompt: > We are looking to pick off one piece of low-hanging fruit based on the state of our roadmap and our codebase. (This will usually be main, but not always.) This will take the shape of a clean up task to improve code quality, a fix for some technical debt you encounter, or something small from our roadmap. > > The roadmap is accessible via our plinky_roadmap MCP server. > > I want you to read through our codebase so you can understand the current state of our app. Then you will read the roadmap and find an item that is not done, but you would estimate has a small level of effort but a high level of return. It's important to note that the two may be slightly out of sync, but that the code is our source of truth. (If there's a completed task on our roadmap that seems to not be marked as complete, you can include that in your response.) > > Then I want you to execute on a code quality improvement, a feature that you think would be valuable to Plinky, or even an improvement to our documentation at an important or tricky point in the code. Every task you do requires a code review so try to not add too much work onto our plate — and focus on high leverage actions. And the results were instantly useful! Codex cleaned up extraneous logging code from debugging Plinky's browser extension, modernized SwiftUI call-sites from the deprecated foregroundColor API to foregroundStyle, and caught that my CSV exporter used \n instead of carriage returns. These are small paper cuts that accumulate over 5 years of work, but fixing them one by one creates a better home for your code. Better yet, Codex went to my roadmap and found a hastily scrawled note that said "Add tests to LinksListFilterStateTests that validate ordering for pinned links." I'd promptly forgotten about this after writing it down — but a few seconds later I had tests. I re-ran the automation a few more times and the results got even better. Codex fixed a bug in how I parse webhook payloads, extracted repetitive code into a private helper function, and found a subtle performance improvement by pre-sorting an array I was re-sorting on every View recalculation. All I did was write a prompt and I had a bug fix, improved code, and a small scrolling hitch had been eliminated. The Future Is Automation It's not hard to see where this is going — any task that fits the shape of software is a task that can be automated. An indie like me is thrilled to have extra hands, but tools like these work for teams small and large. I started with a small scope because code review is still my bottleneck, but I've continued tweaking my prompts and adding more responsibility. Your intern is only as good as your prompts. I've spoken at length about how your ability to prompt and build context for an agentic system is paramount, but I've also set my coding environment up with Codex Skills to make Codex write code the way I like it written. --- Automations are new, so I haven't implemented all of these yet, but here are a few ideas you can try: Nightly issue triage: Scan your issue tracker (Sentry, GitHub Issues, Linear) to assign priorities, identify duplicates, and suggest labels. <details class="callout-toggle"> <summary>Prompt</summary> <div class="callout-body"> Triage open issues for this project. Review open issues from the issue tracker and recent commits for context. For each issue, determine category, relative priority (high/medium/low), and whether it appears to be a duplicate. Produce: 1. A short list of high-priority issues with justification. 2. Suspected duplicates and why they overlap. 3. Suggested labels or metadata. Do not close issues or make destructive changes. </div> </details> Automated architectural docs: Parse code changes and update or infer architectural diagrams and high-level readmes. <details class="callout-toggle"> <summary>Prompt</summary> <div class="callout-body"> Keep architectural documentation aligned with the codebase. Review recent commits, existing README and architecture docs, and overall project structure. Identify meaningful changes to components, data flow, boundaries, or integrations. Propose concise updates to existing documentation, or specific additions where gaps exist. Focus on preserving intent and helping new contributors understand the system. Keep changes minimal. </div> </details> Automated test health monitor: Evaluate your test suite to find flaky tests, suggest missing edge cases, and propose new tests for untested logic. <details class="callout-toggle"> <summary>Prompt</summary> <div class="callout-body"> Evaluate the overall health of the test suite. Review test files, recent failures or flakes, and recently changed code paths. Identify: - Flaky or brittle tests - Important logic paths lacking coverage - Redundant or low-signal tests Produce: 1. A short list of test risks. 2. One or two high-leverage test improvements. Avoid broad rewrites. Optimize for confidence per line of test code. </div> </details> Product surface area audit: Periodically list features users can see, can't discover, and misunderstand. Then suggest one small affordance — a copy tweak, default, or UI nudge to reduce that gap. <details class="callout-toggle"> <summary>Prompt</summary> <div class="callout-body"> I want you to audit the current product surface area of this project. First, read through the codebase, UI components, feature flags, onboarding flows, and any relevant product documentation to understand what the product can do today. Then produce three short lists: 1. Features that are clearly visible and discoverable to users. 2. Features that exist but are difficult to discover without prior knowledge. 3. Features that users are likely to misunderstand based on naming, defaults, or UI affordances. For each item in lists 2 and 3, propose one small, low-risk improvement (copy change, default tweak, UI hint, reordering, or documentation note). Do not propose large redesigns or new features. Optimize for leverage and simplicity. </div> </details> Solicit more ideas: If you had one uninterrupted hour t3b:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaplqoy2f",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/2/12/automating-codex/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering","Productivity"],"$type":"site.standard.document","title":"Automating Codex","description":"You‘re one step away from turning Codex into a proactive assistant that tackles low-hanging fruit while you sleep.","publishedAt":"2026-02-12T12:00:00.000Z","textContent":"$56","canonicalUrl":"https://build.ms/2026/2/12/automating-codex/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaplqoy2f","sort_date":"2026-02-12T12:00:00+00:00"}}] 57:T1ef7,> This post shows you how to support markdown search for your blog, making it token-efficient to search through your content. It also has a lesson about how context windows work and markdown work. If you only want the implementation steps, you can skip all that and go straight to "So What Are We Doing?". If there's one thing LLMs struggle with, it's efficiency. I'm not talking about environmental impact — I'm talking about context windows and tokens. LLMs know a lot of things, but they don't know everything. When they recognize they don't know something, they may fall back to searching the web to learn more. This is great because it means they can access the latest information available to solve the problem. But when LLMs retrieve information from the web, they need to store it somewhere they can reference while working. You can't keep every word of every book you've ever read in your head — and neither can LLMs. Instead, they store retrieved information in a small amount of quickly accessible memory called a context window. A Primer On Context Windows An LLM's context window can't be infinitely large because the data we put in it needs to be stored somewhere, and it has to be accessed quickly. Having to access it quickly rules out traditional long-term storage mechanisms like databases, because they're too slow. In effect, we have two constraints: we want to keep only the most important information, and we need to make it as compact as possible. Before we move on, we need to go over a few facts about context windows: - The text (or images) you put into an LLM are converted into an LLM-friendly format called tokens. - A token is a chunk of data that makes it easier for an LLM to process information. - Counting tokens isn't as straightforward as you'd expect, but here are some rough guidelines: about 3-4 characters of text = 1 token, or roughly 1 word = 1.33 tokens. - The average Harry Potter book is 135,000 words long, or about 180,000 tokens. - The latest models at the time of this writing are Claude Opus 4.5 and GPT-5.2 — which have context windows of 200K and 400K tokens respectively. - This is a pretty limited space, so once you've reached the end of your token window, LLMs will do their best to summarize the conversation. Summarization, though, can lose important details. This means you can safely fit 1-2 Harry Potter books in a context window — which is far smaller than the entirety of the internet. In practice, that's approximately 10-15 of my blog posts, or roughly 2-3 hours of video on YouTube. And that's why we need to be selective when we ask LLMs to search the web for information. When Claude goes off to search the web, it may retrieve far more data than its context window can hold! And That's Why We Need Markdown The good news is that my blog posts are written in markdown — a lightweight markup language created by John Gruber. If you're reading this, you've likely used markdown to bold or italicize text by adding _ or * around words or sentences. Markdown is very lightweight compared to HTML because it's plain-text without any markup. (Hence the name markdown — which I only just realized as I was writing this sentence. 😆) Sharing the markdown version of a blog post will save precious tokens — which lets LLMs store and access more information in their context window. <blockquote class="twitter-tweet" data-dnt="true" data-theme="light"><p lang="en" dir="ltr">In the next version of Claude Code, Claude’s WebFetch tool automatically adds Accept: “text/markdown, ” to requests which helps docs sites provide token-efficient docs <a href="https://t.co/CZWdH0OZsO">https://t.co/CZWdH0OZsO</a> <a href="https://t.co/uYthd1m9RP">pic.twitter.com/uYthd1m9RP</a></p>— Boris Cherny (@bcherny) <a href="https://twitter.com/bcherny/status/1988860326306087102?ref_src=twsrc%5Etfw">November 13, 2025</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script> This approach has become standard practice — Claude Code now automatically searches for markdown versions of websites, and other tools like Codex have followed suit. Yet despite its importance, few websites have implemented this technique. Here's how to be one of them. So What Are We Doing? We're going to do two things. 1. Add support for serving markdown versions of our blog posts. 2. Direct LLMs to browse the markdown version instead of our HTML-rendered blog posts. Or more technically: 1. When someone adds .md to a blog post URL, we'll serve the markdown version of that post. (Try it yourself with this post: build.ms/2026/2/2/supporting-markdown-search-for-llms.md) 2. We'll build a small middleware layer that redirects LLMs to the markdown version when they request a page with an Accept header of text/markdown. Adding Markdown Support To Your Website Build.ms uses the Astro static site generator, but this technique works for any static site generator. To give you an idea of what we want to do, here's the prompt I gave to Codex to build this feature: > Is it possible for us to expose the original markdown of our posts whenever the user adds a .md extension to a URL? This has become a common pattern for letting AI tools crawl websites, providing a more direct format that they can consume easily. The result was a file called [slug].md.ts. This file convention creates a route that serves the markdown version of your blog post. Now when a person (or LLM) adds .md to any blog post URL on build.ms, they see the markdown version of our post. <details class="callout-toggle"> <summary>Expand to see the code for [slug].md.ts</summary> <div class="callout-body"> </div> </details> We can now start dropping links with .md extensions in our prompts and watch our token count drop. But it would be better if this process was automatic. To do that, I wrote another prompt that explains our goals, so Codex can solve our problem. > I just want to check something. The reason we built this was because I saw this post from the Claude Code team. > \> In the next version of Claude Code, Claude’s WebFetch tool automatically adds Accept: “text/markdown, ” to requests which helps docs sites provide token-efficient docs > The reason we recently added support for .md extensions rendering markdown docs was to support this feature, but are we handling that? If not, can we add that somehow? <details class="callout-toggle"> <summary>Expand to see the code for our middleware</summary> <div class="callout-body"> </div> </details> And just like that, Codex added a middleware layer that redirects any request with an Accept header of text/markdown to the .md version of our blog post. Now the two-step process is complete, and our token usage will automatically be optimized. And That's It? As you can see adding support for the text/markdown Accept header is straightforward and easy! And it's beneficial: LLMs get more accurate information, they use fewer tokens, and your website saves on bandwidth costs. You can implement this feature yourself, but rather than writing the code, let's use the technique I shared in Building Software From Blog Posts. Point Codex or Claude Code to this blog post and let it handle the implementation. Here's a prompt you can use — with all the context* you'll need. > Please read this blog post (https://build.ms/2026/2/2/supporting-markdown-search-for-llms) that describes a process for adding support for the text/markdown Accept header — allowing LLMs to more easily and efficiently parse the contents of a statically generated blog. We want to implement the same technique for our static site generator — regardless of whether it's Astro or not. If there are any concerns please raise them before we begin building this feature. If there are no concerns, please implement a solution that solves this problem for us in our static site.3c:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapqqmr2a",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/2/2/supporting-markdown-search-for-llms/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Tutorial","Engineering"],"$type":"site.standard.document","title":"Supporting Markdown Search For LLMs","description":"Learn a simple but effective technique that makes your blog more easily searchable for LLMs.","publishedAt":"2026-02-02T12:00:00.000Z","textContent":"$57","canonicalUrl":"https://build.ms/2026/2/2/supporting-markdown-search-for-llms/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapqqmr2a","sort_date":"2026-02-02T12:00:00+00:00"}}] 58:Ta2c,Generative Engine Optimization — also called AI SEO, AEO, GEO, or LLMO — is the equivalent of SEO for AI search. Regardless of what acronym you prefer, it's becoming increasingly important for agentic tools like ChatGPT, Claude, and Gemini to know about you and your website — whether you're a developer, marketer, content creator, or business owner. I've never been particularly good at SEO, so I was very excited when I stumbled upon this post by Cassidy Williams. To quote Cassidy Williams (a brilliant developer, writer, and content creator), the most important things you can do are: > - Creating an /llms.txt file > - Adding “LLM-readable” structured pages (often called /for-llms) > - Adding Schema.org data to the pages of my website > - Being consistent with naming, phrases, and taglines Naturally, instead of doing all of this myself, I wrote this prompt in Codex and had it generate all of these files for me! > I want to update my blog to be more generative AI search-friendly. > > 1. Can you read this blog post to understand how we can do this: https://cassidoo.co/post/ai-llm-discoverability > 2. I want you to then read through my src/pages/about.astro page, knowing that a lot has changed about me since I wrote it. I want you to read all of the blog posts we have on this website and search the web and to find everything you can about Joe Fabisevich, then lightly update the page accordingly based on my goals and mission for the website. Keep the writing style the same though so we don't lose my voice. > 3. As per the blog post then I want you to write a great llms.txt, for-llms page, and anything else that’s missing for AI tools to know about build.ms. > 4. I noticed we're missing a robots.txt which helps AI crawlers know how to browse my site, so make sure to add one. > > Before we start, do you have any clarifying questions I can answer to make this process work better? I answered a few clarifying questions Codex had about my website and did some light editing to keep my voice intact — and that was it. Codex did 98-99% of the work, and while it was generating those files, I'd already started writing this blog post. --- It's more important than ever to be discoverable on the internet, and GEO is the next generation of discoverability. If you have a website and want to be found, you'll need to make your website GEO-friendly. The good news is that agentic tools are great at generating GEO-friendly pages. All you need to do is copy the prompt above — change a few words — and let Codex do the rest. Can't wait to find you on the internet!3d:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapt7fn2a",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/notes/2026/1/29/updating-your-website-for-generative-engine-optimization/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Marketing","Design","Engineering","Business","SEO","GEO"],"$type":"site.standard.document","title":"Updating Your Website For Generative Engine Optimization","description":"Learn how to optimize your website for generative search engines like ChatGPT, Claude, and Gemini to improve your search engine rankings.","publishedAt":"2026-01-29T12:00:00.000Z","textContent":"$58","canonicalUrl":"https://build.ms/notes/2026/1/29/updating-your-website-for-generative-engine-optimization/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapt7fn2a","sort_date":"2026-01-29T12:00:00+00:00"}}] 59:T2097,A few weeks ago I read an essay titled AI Agents Are Starting To Eat SaaS. SaaS businesses have thrived by solving specific problems — outsourcing domain-specific knowledge so companies don't have to build it themselves. A software company buys Workday instead of building HR software. A healthcare company uses Salesforce instead of maintaining a custom CRM. A piano teacher pays $9.99/month for Calendly instead of building their own booking system. The responses to that essay were full of skepticism — people saying they haven't noticed any real decline in SaaS adoption. That may be true for them, but the most important word in the title is Starting. It's too early to see a trend, but I've already heard dozens of stories where people are building internal solutions through prompting rather than turning to a SaaS tool. I've done this myself quite a few times since I'd rather spend 30 minutes building something personalized to my needs than pay $4.99 every month for a generic solution. --- <blockquote class="bluesky-embed" data-bluesky-uri="at://did:plc:b6eke66r3vbmnegg73qgprl6/app.bsky.feed.post/3mc3hq477jk2n" data-bluesky-cid="bafyreiablauwd6dkws3glacx7nqzg7aw6qeeyrg5y2vfezvp5uqc5fimr4" data-bluesky-embed-color-mode="system"><p lang="en">I was wondering how long exactly it's been since I first articulated this exact scenario, and it turns out the answer is 33 months.<br><br><a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6/post/3mc3hq477jk2n?ref_src=embed">[image or embed]</a></p>— Joe Fabisevich (<a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6?ref_src=embed">@mergesort.me</a>) <a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6/post/3mc3hq477jk2n?ref_src=embed">January 10, 2026 at 11:49 AM</a></blockquote><script async src="https://embed.bsky.app/static/embed.js" charset="utf-8"></script> I've been called the boy who cries wolf for three years, predicting that soon enough building your own app would become cheaper and easier than buying SaaS — and that custom solutions would be better because they're tailored to your needs. The main pushback has always been that it requires skill to prompt effectively, and most people lack that know-how. But LLMs keep improving while AI tools keep getting easier to use. This is really important because these improvements compound, lowering the barrier to entry for millions of people. <blockquote class="bluesky-embed" data-bluesky-uri="at://did:plc:b6eke66r3vbmnegg73qgprl6/app.bsky.feed.post/3mc3eq6pnks2u" data-bluesky-cid="bafyreih26ass24xwfj3mrsgqf2uvade5fm5oqmdf6lpx2e6dmloetusyci" data-bluesky-embed-color-mode="system"><p lang="en">As a software developer who spends his days building a link-saving app and his nights telling everyone that in the very near future people will be able to easily build their own software to solve their personal problems — this is my existential crisis coming to life. <br><br><a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6/post/3mc3eq6pnks2u?ref_src=embed">[image or embed]</a></p>— Joe Fabisevich (<a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6?ref_src=embed">@mergesort.me</a>) <a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6/post/3mc3eq6pnks2u?ref_src=embed">January 10, 2026 at 10:55 AM</a></blockquote><script async src="https://embed.bsky.app/static/embed.js" charset="utf-8"></script> As the boy who cried wolf, I feel confident that a wolf has finally walked into a town full of software developers building SaaS products. The person who wrote the thread I'm quoting isn't a software developer — it's Kevin Roose, a reporter for The NY Times. He missed Pocket after it shut down, so he tried every link-saving app available. None met his needs, so over the Christmas break he built his own replacement in 12 prompts. It worked perfectly for him — so well that people begged him to open-source it. As the developer of Plinky — another link-saving app — I'm watching this moment like a dinosaur looking up at the sky wondering about the impending meteor. <blockquote class="text-post-media" data-text-post-permalink="https://www.threads.com/@crumbler/post/DTRX5P7ie3w" data-text-post-version="0" id="ig-tp-DTRX5P7ie3w" style=" background:#FFF; border-width: 1px; border-style: solid; border-color: #00000026; border-radius: 16px; max-width:650px; margin: 1px; min-width:270px; padding:0; width:99.375%; width:-webkit-calc(100% - 2px); width:calc(100% - 2px);"> <a href="https://www.threads.com/@crumbler/post/DTRX5P7ie3w" style=" background:#FFFFFF; line-height:0; padding:0 0; text-align:center; text-decoration:none; width:100%; font-family: -apple-system, BlinkMacSystemFont, sans-serif;" target="_blank"> <div style=" padding: 40px; display: flex; flex-direction: column; align-items: center;"><div style=" display:block; height:32px; width:32px; padding-bottom:20px;"> <svg aria-label="Threads" height="32px" role="img" viewBox="0 0 192 192" width="32px" xmlns="http://www.w3.org/2000/svg"> <path d="M141.537 88.9883C140.71 88.5919 139.87 88.2104 139.019 87.8451C137.537 60.5382 122.616 44.905 97.5619 44.745C97.4484 44.7443 97.3355 44.7443 97.222 44.7443C82.2364 44.7443 69.7731 51.1409 62.102 62.7807L75.881 72.2328C81.6116 63.5383 90.6052 61.6848 97.2286 61.6848C97.3051 61.6848 97.3819 61.6848 97.4576 61.6855C105.707 61.7381 111.932 64.1366 115.961 68.814C118.893 72.2193 120.854 76.925 121.825 82.8638C114.511 81.6207 106.601 81.2385 98.145 81.7233C74.3247 83.0954 59.0111 96.9879 60.0396 116.292C60.5615 126.084 65.4397 134.508 73.775 140.011C80.8224 144.663 89.899 146.938 99.3323 146.423C111.79 145.74 121.563 140.987 128.381 132.296C133.559 125.696 136.834 117.143 138.28 106.366C144.217 109.949 148.617 114.664 151.047 120.332C155.179 129.967 155.42 145.8 142.501 158.708C131.182 170.016 117.576 174.908 97.0135 175.059C74.2042 174.89 56.9538 167.575 45.7381 153.317C35.2355 139.966 29.8077 120.682 29.6052 96C29.8077 71.3178 35.2355 52.0336 45.7381 38.6827C56.9538 24.4249 74.2039 17.11 97.0132 16.9405C119.988 17.1113 137.539 24.4614 149.184 38.788C154.894 45.8136 159.199 54.6488 162.037 64.9503L178.184 60.6422C174.744 47.9622 169.331 37.0357 161.965 27.974C147.036 9.60668 125.202 0.195148 97.0695 0H96.9569C68.8816 0.19447 47.2921 9.6418 32.7883 28.0793C19.8819 44.4864 13.2244 67.3157 13.0007 95.9325L13 96L13.0007 96.0675C13.2244 124.684 19.8819 147.514 32.7883 163.921C47.2921 182.358 68.8816 191.806 96.9569 192H97.0695C122.03 191.827 139.624 185.292 154.118 170.811C173.081 151.866 172.51 128.119 166.26 113.541C161.776 103.087 153.227 94.5962 141.537 88.9883ZM98.4405 129.507C88.0005 130.095 77.1544 125.409 76.6196 115.372C76.2232 107.93 81.9158 99.626 99.0812 98.6368C101.047 98.5234 102.976 98.468 104.871 98.468C111.106 98.468 116.939 99.0737 122.242 100.233C120.264 124.935 108.662 128.946 98.4405 129.507Z" /></svg></div><div style=" font-size: 15px; line-height: 21px; color: #000000; font-weight: 600; "> View on Threads</div></div></a></blockquote> <script async src="https://www.threads.com/embed.js"></script> Now I know that not everyone will build their own app. My mother-in-law certainly won't build her own Plinky-replacement and will happily keep supporting her son-in-law's work. But for companies like Squarespace, reporters vibe-coding their own websites in just a few hours rather than paying $200/year ought to be a red-alert moment. This won't affect every SaaS product — enterprises rely on SaaS tools for stability and dependability that you may not want to in-source — but AI will drastically chip away at many businesses. --- I've written several posts about how on-demand software is already here and that people are already capable of building their own solutions even if not everyone knows it yet. More and more people are catching on, and tools like Claude Code are evolving into new modalities like Claude Cowork that let any knowledge worker access the power developers have had for most of 2025. The changes from this shift will be drastic and will disrupt many professions and industries. How this plays out deserves deeper exploration, but one thing is clear: it will upend many SaaS businesses. It's no longer a question of if, but when. Especially for companies that don't adapt to this new reality.3e:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapvqxq2z",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/1/26/ai-agents-are-starting-to-eat-saas-really/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Business","AI Economics"],"$type":"site.standard.document","title":"AI Agents Are Starting To Eat SaaS (Really)","description":"AI is making it cheaper and easier for anyone to build custom software solutions. The changes from this shift will be drastic, especially for SaaS companies.","publishedAt":"2026-01-26T12:00:00.000Z","textContent":"$59","canonicalUrl":"https://build.ms/2026/1/26/ai-agents-are-starting-to-eat-saas-really/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqapvqxq2z","sort_date":"2026-01-26T12:00:00+00:00"}}] 5a:Tecc,<blockquote class="bluesky-embed" data-bluesky-uri="at://did:plc:b6eke66r3vbmnegg73qgprl6/app.bsky.feed.post/3mcij2xhtgc2x" data-bluesky-cid="bafyreigm5qpdgpqnsadjq5dur5535njdmllcubmwwlqqqno2gpddbls2o4" data-bluesky-embed-color-mode="system"><p lang="en">I feel like we don’t talk enough about how five years ago the only people doing interesting things with “AI” were researchers and big tech employees optimizing algorithms, and now everybody has powerful AI systems available to them on demand via an API or in practically every tool.</p>— Joe Fabisevich (<a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6?ref_src=embed">@mergesort.me</a>) <a href="https://bsky.app/profile/did:plc:b6eke66r3vbmnegg73qgprl6/post/3mcij2xhtgc2x?ref_src=embed">January 15, 2026 at 4:17 PM</a></blockquote><script async src="https://embed.bsky.app/static/embed.js" charset="utf-8"></script> --- When I worked at Twitter, I worked across many teams and was involved in features across the platform. There was one thing I could never wrap my head around though — the AI[^1] that powered the Home timeline. I could feel when the quality of my feed dropped or things felt worse, but it was all qualitative and I couldn't explain why. I simply didn't have the expertise to explain why, nor the skills to do anything about it. Today you don't have to understand neural nets and tensors to build something with AI. With just an API key, everyone has access to OpenAI and Anthropic's models and can do amazing things with AI in a way that simply wasn't possible five years ago. Heck, the barrier to entry is even lower if you consider more no-code interfaces like n8n, Make, and Zapier. Now you have to understand how to work with AI, but previously AI was this mystical domain that was accessible to practically nobody. As I discussed in The Claude Code Moment, you don't have to be an AI researcher anymore — you can play, experiment, and prompt your way to making something interesting. It no longer takes weeks of training a model to try something new, all you need is curiosity. You don't have to tweak parameters or understand neural networks — you just update to the latest version of Claude or GPT-5 and see your results improve. Judging by how many people are building things with AI today versus a few years ago, the barrier to entry has dropped thousands-fold. This shift from AI being accessible to a few people in labs to anyone on the planet is extraordinary. You no longer need a PhD or a job at Meta or Google to use powerful AI systems to build something valuable, unique, and meaningful. All you need is an internet connection, plus some curiosity, creativity, and critical thinking skills. I have a core belief that people are ingenious. Given the tools to implement their ideas, there's no telling what they'll create. Democratizing access to AI isn't just about lowering the barriers — it's about unleashing human creativity at scale — and I find that beautiful and exciting. We're living through a historical moment, but we may not even recognize it as such because we're in the middle of it. Five years ago, practically nobody would have predicted how exponentially AI would penetrate every corner of our lives — and it's just as hard to predict what the landscape will look like five years from now. People talk about the ChatGPT moment or the more recent Claude Code moment as turning points — but only in hindsight because you must live through a moment to truly understand its impact. Right now, we're living through a moment where AI has become accessible to everyone, and we're only beginning to see what becomes possible. [^1]: Admittedly back then we called it machine learning or deep learning, but it was the bleeding edge of AI.3f:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaq2m3s2p",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/1/21/the-democratization-of-ai/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","AI Economics"],"$type":"site.standard.document","title":"The Democratization of AI","description":"The barrier to building with AI has dropped thousands-fold. What happens when you give everyone the tools to implement their ideas?","publishedAt":"2026-01-21T12:00:00.000Z","textContent":"$5a","canonicalUrl":"https://build.ms/2026/1/21/the-democratization-of-ai/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaq2m3s2p","sort_date":"2026-01-21T12:00:00+00:00"}}] 5b:T12f8,For the last year I’ve been teaching people in Pursuit’s AI-Native. As I described in Build.ms: Making Technology Accessible To Everyone > I've been volunteering at Pursuit for 12 years — helping people from below the poverty line, with no prior background in tech, become software developers. 80% of them land jobs in tech making an average salary of $85,000/year. Former Uber drivers, janitors, and kindergarten teachers have learned technology skills to completely reshape their lives. I've had the great pleasure of watching hundreds of lives change right in front of my eyes, and it's truly moving. > Given the poor job market for junior developers, the program shifted to an AI-focused approach last year. People who've never heard of AI are building real software and learning how to build real products in just 9 months. They put in hundreds of hours learning to build with AI, develop products, market them, and everything else you'd need to build a small business. They do this with guidance from dedicated teachers and volunteers who transform people into builders incredibly quickly. These are incredibly dedicated people learning everything they need to thrive in an AI-centric world. They saw the opportunity a year ago and decided to jump in before most people had even heard of Claude Code. As I wrote in Claude Cowork > Since the release of Claude Code, software developers have been living in the future. As Simon Willison puts it, developers have had access to general purpose agents for the last 11 months. But Claude Code is finally having its moment with non-technical people who are realizing they don't need to be technical to use it to write code. Many had assumed it would be intimidating or require a coding background, but are now discovering that AI can guide them through the parts where they'd normally get stuck. --- The Hard Fork podcast asked listeners to share how they're using Claude Code — specifically regular people, not developers. (It starts at 36:07 in the episode if you want to listen.) There were stories like: - A woman who loves to read built a service where people can sign up for personal book recommendations — just for fun! - A person whose business is selling wallpaper to interior decorators built her own wallpaper calculator. Decorators can now submit their wall sizes and get exactly how much wallpaper they need to purchase, plus a layout diagram. In her words: this was always a job I used to do by hand in Adobe Photoshop or InDesign, but now clients can do it themselves and save time for us both. - A man who runs a CNC milling business built agents that generate leads for jobs in his local area, and even created his own MCP server to interface with his CAD software. He also wrote a personal message to the podcast hosts: > I really want people in my demographic to understand that this isn't “I asked ChatGPT to write an email or to generate an image that has the correct spelling.” This is different to me, I built my own business infrastructure from scratch with an AI pair programmer, despite having zero formal training and a high school education. This is exactly what I've been seeing at Pursuit for the last year. People who learn to use agentic tools like Claude Code become more empowered, self-sufficient, and gain newfound autonomy to make what they need — without the technical background that was previously required. The skill of commanding an agentic system to build something meaningful has little to do with writing code — all it requires is clarity of thought, creativity, and critical thinking skills. This is wonderful for non-technical people — the power to create software has become dramatically more accessible. And for developers, this is worth paying attention to. People are crossing into technical domains without waiting for permission or credentials now that they can build solutions to their own problems. If you've considered writing code your moat, it's worth reconsidering how quickly that's going to change across a wide swath of domains. --- When I was starting my career, it was common to see people list Microsoft Word, Excel, and PowerPoint on their resume. Today, everyone assumes you can use those (or their Google equivalents) if you're applying for a knowledge work job. The ability to use AI well will become the same — something your employer assumes you know how to do. Not everyone will be equally skilled at it, the same way some people are wizards with a spreadsheet while others use it as a glorified table. That's why I teach AI workshops — because I genuinely believe this is a key skill for the future of work — and I've helped unlock something in hundreds of people. And that's why if you haven't tried Claude Code yet — or its less technical sibling Claude Cowork — this is your moment.40:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaq53sb2d",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/1/19/the-claude-code-moment/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Society","AI Economics"],"$type":"site.standard.document","title":"The Claude Code Moment","description":"The sudden adoption of Claude Code by non-technical people represents an important moment in time — and one that will have a lasting impact on the way we work.","publishedAt":"2026-01-19T12:00:00.000Z","textContent":"$5b","canonicalUrl":"https://build.ms/2026/1/19/the-claude-code-moment/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaq53sb2d","sort_date":"2026-01-19T12:00:00+00:00"}}] 5c:Tfdd,Build.ms exists to make technology accessible to everyone. The technological landscape is moving faster than ever. Even professional developers, designers, and product managers are struggling to keep up with every new advancement. (And I'm not just talking about AI.) All the noise makes it hard to filter out what's important, and what you can ignore. And people have busy lives. There's work to do, families to raise, friends to see, and hobbies to pursue. You shouldn't have to spend all your time keeping up with the daily flood of news to understand how technology is reshaping the world around you. Build.ms has always been a space for technical writing to share lessons about engineering, product development, design, marketing, AI, and anything you need to build a modern business. But I'm reshaping it around a clearer purpose: making technology accessible to everyone. --- This isn't a new passion — I've been living this mission for over a decade, but I now have the clarity to make it my primary goal. I've broken complicated subjects down into simple step-by-step processes so you can learn by doing. I've demonstrated new ways of working that are emerging. I've given talks that teach new ways to think about technology. Each of these is designed to bridge the gap: to give beginners a way in while offering experts new clarity. That's what you'll find here going forward: more frequent posts that break down complex subjects into something you can understand and use immediately. --- I've been volunteering at Pursuit for 12 years — helping people from below the poverty line, with no prior background in tech, become software developers. 80% of them land jobs in tech making an average salary of $85,000/year. Former Uber drivers, janitors, and kindergarten teachers have learned technology skills to completely reshape their lives. I've had the great pleasure of watching hundreds of lives change right in front of my eyes, and it's truly moving. Given the poor job market for junior developers, the program shifted to an AI-focused approach last year. People who've never heard of AI are building real software and learning how to build real products in just 9 months. They put in hundreds of hours learning to build with AI, develop products, market them, and everything else you'd need to build a small business. They do this with guidance from dedicated teachers and volunteers who transform people into builders incredibly quickly. Volunteers like me spend hours every week filling in the gaps, and we hear how valuable it is. But right now, those lessons are only available when I share an insight on Bluesky, answer a question in the communities I'm a part of, or teach a hands-on workshop. The knowledge is there, but it's not accessible to everyone who needs it. That's not enough. This information needs to be free and accessible — to everyone. --- Here's what I've learned from a decade of teaching: The people who thrive aren't the ones who know the most. The people who thrive are the ones who know how and what to learn, how to adapt to new information, and how to think critically and creatively. There are a million places to learn about a specific tool or technology, but without the right mental models — taught in an accessible way — most people never figure out how to use those tools effectively. That's my mission for Build.ms 2.0. It's a subtle shift from giving you everything you need to build a modern business, to sharing what matters most in a fast-moving technical world. I want people to have the knowledge they need, and to do something good with it. Whether that's making your life a little easier, helping the people around you, or paying it forward — it all starts with information. I hope I can help you with that, and I'd love for you to follow along with me. > Email Newsletter: build.ms/newsletter > > Bluesky: bsky.app/profile/mergesort.me > > Mastodon: macaw.social/@mergesort > > Threads: threads.com/mergesort > > LinkedIn: linkedin.com/in/devjoe > > RSS: build.ms/feed.xml41:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqgtdf2a",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2026/1/6/making-technology-accessible-to-everyone/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Programming","Industry","Productivity"],"$type":"site.standard.document","title":"Build.ms: Making Technology Accessible To Everyone","description":"Learn how to think about technology in a fast-moving world — I‘ll show you what matters and filter out the rest.","publishedAt":"2026-01-06T12:00:00.000Z","textContent":"$5c","canonicalUrl":"https://build.ms/2026/1/6/making-technology-accessible-to-everyone/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqgtdf2a","sort_date":"2026-01-06T12:00:00+00:00"}}] 5d:T271b,It's a Christmas surprise! My AI (Without the Hype) talk from NSSpain is now available for everyone to watch. This talk is wide-ranging and covers a lot of ground. I start by showing you a day in the life of using AI, discuss the fundamentals of how AI systems work, and then share ways to shape the way you work with AI to get great results. Whether you're a beginner or an expert, a developer or a designer, or just someone who wants to learn more advanced ways of thinking about AI — there's something for everyone. I'm not saying you should ignore your family on Christmas Eve, but if you're looking to learn about a lot about AI once the mice are no longer stirring about — I think you'll really enjoy my talk.  <iframe width="720" height="404" src="https://www.youtube.com/embed/I-SiXRYHfEM?list=PLztE34GS_piKKQ6y1dkkuhW76jLBHm3NV" title="AI (Without the Hype) - Joe Fabisevich" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen> - Here's a direct link - Better yet, you can save it to my app Plinky, and watch it later. - And here's the full transcript below. It was pulled from YouTube and cleaned up with Claude Code, but I did skim it over to make a few minor edits and didn't notice anything that misquoted me. --- Introduction Daniel Steinberg: I know I talk about how old I am a lot, but one of the reasons is people ask me why I don't retire. And the reason I don't retire is because of this. I mean, most of my friends are elsewhere. Daniel Steinberg: So, I know I don't remember names, but I really appreciate talking to people. And on the bus ride up from Madrid to Logroño, which is quite a bus ride, Joe said, "Hey, mind if I sit up here? We chat?" And it's a great way to meet somebody that you only kind of know online. And people online, he's much nicer in person. And I'll tell you that he's an indie iOS developer. You might know him from his app Plinky. He also teaches an AI workshop because that's what everyone does these days. He's going to talk to you about AI without the hype. Please welcome Joe. --- Who Am I? Thank you so much for having me here at NSSpain. It's truly an honor to be here at this wonderful conference with so many of my favorite iOS developers. As you can guess, I'm here to talk to you about AI. But first, I'd like to tell you a little bit about who I am. I'm an indie developer and I spend my days working on an app called Plinky. It's the easiest way to save links for later, available on the App Store, and people really seem to like it. I've worked at every size company from a startup of two to internet success stories like Time Hop and Bitly. Most recently, I spent four years helping build the health client team at Twitter when it was Twitter, with a focus on improving misinformation and abuse at a global scale. I'm an open source maintainer and a writer, and these days I'm also a teacher. I help people of all backgrounds, technical and non-technical, learn all about AI, as we'll be doing here today. I love teaching people about this technology that's emerging in real time. So, I work with individuals and teams alike to develop highly personalized hands-on workshops. What I Love About Code A simple and reusable abstraction tickles my brain, as does the personal fulfillment of solving a really difficult challenge. Code on its own is interesting, but I also love using code to build products that people love. Don't get me wrong, as an indie, I have to do design, marketing, sales, customer support, and of course, build my product. I love doing each of these things a little less than I love to code. Some, a lot less, actually. But what I love most is making a real connection with another person through the software that I build. I feel so lucky that I get to build personal, playful productivity apps that make people happy. I'm always moved when a person reaches out to tell me, and that's what drives me to do all the things that I don't necessarily want to do. These are the roles that I take on to enable me to keep doing what I love most. And that's why I'm going to be talking to you today about how to use AI for your job as a software developer to do more of the things that you love most and less of the things that you don't. What We'll Cover We'll start by talking about where AI is today and how I use AI in my day-to-day life. Then we'll talk about the technology behind AI and why AI tools work the way that they do. This will help explain why AI will continue to improve and diffuse across the industry, reshaping the way that we all work. Most importantly though, we'll discuss the techniques that are required for successfully working with AI to accomplish your goals. There's a lot that AI can do today and plenty that it can't. I'm not here to convince you that it's a magical tool that can do anything and everything. And that's why I have this disclaimer. A Disclaimer As I mentioned, this talk is about AI, and people have a lot of feelings about AI — understandably so. Because I help people learn about AI, I hear all of these feelings all of the time. Some people think that AI can't do very much and that it's just a lying hallucination machine. Others worship at the altar of Sam Altman and think that we're building a digital god. Much like everything in life, I believe the answer is somewhere in the middle. There are many moral, societal, and security concerns to be had about AI — there are so many of them — and I genuinely wish that we were thinking deeply when having those discussions. The field is changing very quickly, so a few of the things that I present won't be the only approach or the only solution. I personally use AI a lot, and I'll walk you through how it's been a multiplier for my productivity. While I get a lot of utility out of AI, I definitely do not delegate all of my thinking to it. What remains most important as a software developer and as a person is your ability to think critically and to adapt to new technologies. That's why I want to ask that for the next 30 minutes, you put aside your preconceived notions about AI, regardless of what they might be. And then I'd love for you to keep an open mind to the ideas that I present. In 31 minutes, you can believe whatever you like. I promise. --- A Day in My Life with AI I want to start by walking you through a regular day in my life, specifically the ways that I use AI. I start most mornings with a 15-minute bike ride up and down a quiet street where I live in New York City. Sometimes I listen to a podcast. Sometimes I enjoy the sounds of the city. But lately, I've been opening ChatGPT's voice mode and simply saying, "Teach me something new." Learning with AI It's like having an encyclopedia on demand. If I don't know something, I just ask. If I find something interesting, I dive deeper down the rabbit hole. And of course, if I'm skeptical about something, I follow through and get more clarity. LLMs are like lossy encyclopedias. The more factual information there is about something in the training data, the more accurate you can expect an LLM to be. There's a lot of highly sourced information about bioluminescence. So, I feel pretty good about the validity of what I've learned. You can learn about almost anything. If it's available on the internet, you can probably learn about it. It could be science, it could be history, or it can even be code. Lately, my bike rides have been filled with question after question about recent Swift evolution proposals. I'm not looking for an easy substitute for learning. Building a deep understanding requires practice and hard work. Instead, I'm using a new approach to break down a dense subject by asking the questions that I need answers to. Building Tools on Demand As I'm walking home from my bike ride, I have an idea. I get back to working on my NSSpain talk and I use AI to build this slide. Every slide in this talk with a chat bubble was made with a little tool that I built in a Claude artifact. The chat bubbles were created to be an exact replica of ChatGPT but with my talk's color scheme. It was so simple that I built it with the prompts that you see on screen in less than 5 minutes. I didn't have to write any code. I just had to describe my problem and give that to Claude. Artifacts are accessible via the web, so you can go play with mine or even create your own based on it. Now, is this chat bubble generator my life's work? Of course not. It's just a simple tool. But humans are tool builders. Being able to build tools on demand with AI is what makes AI so empowering to a software developer. This XKCD comic about automation used to be a joke, but now it's our reality. Rather than sitting in Keynote all day manipulating shapes until they look just right, I can now automate the process of creating a piece of software. Three prompts and five minutes later, I had exactly what I needed. No code and no thinking required to create a usable and reusable piece of software. Many problems are harder to describe than this, but sometimes it's actually this easy to describe and build something that you need. The cost of making throwaway software that helps you accomplish a specific task is now almost zero. Natural Language API Testing Take a look at these two APIs and tell me which one you like more. In the spirit of automation, halfway through a curl request to test Plinky's API, I realized I'm sick and tired of writing curl requests. I could use a GUI tool designed for API requests like Paw or Postman. But it would be even better to have a personalized tool. What if I could simply type "create tag NSSpain" and AI would create the correct curl request for me? I think about how I can do this and I have a brilliant idea. I open up Codex, which is OpenAI's CLI tool similar to Claude Code. First I ask Codex42:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqjek62s",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2025/12/24/ai-without-the-hype-ns-spain-2025/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Conference Talk","iOS","Programming"],"$type":"site.standard.document","title":"AI (Without the Hype)","description":"I spoke at NSSpain this year and gave a wide-ranging talk AI, building from the fundamentals to real-world uses — with something for everyone.","publishedAt":"2025-12-24T12:00:00.000Z","textContent":"$5d","canonicalUrl":"https://build.ms/2025/12/24/ai-without-the-hype-ns-spain-2025/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqjek62s","sort_date":"2025-12-24T12:00:00+00:00"}}] 5e:T178c,Every programmer has their favorite language. Some prefer Python, while others swear by TypeScript[^1]. Many teams build their apps on Postgres, while others use MySQL. These choices are often flame-war bait for programmers with strong opinions, but most of these decisions are centered around pragmatism, priorities, and tradeoffs. They're reflections of different working styles, rather than moral statements. All of this applies to the choices people make when they decide whether to use Claude Code or Codex. Before we continue, I need to make a disclaimer: This post is about the Claude Code and Codex, on December 22, 2025. Everything in AI changes so fast that I have almost no expectations about the validity of these statements in a year, or probably even 3-6 months from now. I also must emphasize that both Codex and Claude Code are already superhuman developers. I don't say that solely based on the quality of outputs from Opus 4.5 and codex-5.2-high, but in how they work. Codex and Claude Code sometimes arrive at a solution in ways that almost feel alien to how we think about coding, much like AlphaGo's Move 37. --- I keep up with every AI tool I can so I can teach AI to anyone and everyone, but it's becoming more of a necessity for all software developers. When it comes to coding I've mostly settled into using Codex for "coding", and I put coding in quotes because the process is very different than writing code by hand. I spend anywhere from 30 minutes to two hours writing prompts and generating context for Codex, then the task runs for 15-20 minutes while I context-switch to something else entirely. When I come back, I've got somewhere between a day and a week's worth of code waiting for me. But I still use Claude Code a lot. The coding environment they've built is exceptional. Much like Peter Steinberger describes Claude Code as his computer, I delegate all sorts of tasks to Claude Code. I'll use Skills to transcribe videos to mp3s, generate a dark mode color palette for my website, or quickly prototype an idea from a blog post. Why People Love Codex When I need something done and done right, I call on Codex. The reason is simple: the results are unbelievably good for me. And that's because of how I work. By investing time into context engineering and context plumbing, I'm able to stay hands-off-keyboard for much longer. Having long-running tasks might sound like a drawback, but it's just a different way to work. When I send Codex off to do a task that takes 20 minutes, I switch my focus entirely. I'll open Figma to do some design work, write my newsletter, or open another terminal and prompt Codex with some server work while the first terminal is chugging along on some client work. I'd rather have everything take longer and generate results that I don't have to fix than be involved in the process steering AI to success. That's the working style that suits me best. OpenAI's latest models deliver better results than the latest Claude models with less need to be in the loop. So while that remains true, I’ll mostly stick with Codex. Why People Love Claude Code And that brings me to my theory of why many engineers prefer Claude Code. Claude makes you feel more like you're doing engineering work — and surprise — engineers love engineering work. Source: I'm an engineer. Claude has a lot of knobs to turn. You've got your CLAUDE.md, Skills, Agents, MCP, slash commands, and so much more. Codex has similar features, but it tends to produce high-quality results out of the box. Claude on the other hand works best when you finely tune those knobs, and Anthropic really encourages you to do so in their developer relations and marketing. This is a perfect match for engineers who love configuring their environments. I can't tell you how many full days of my life I've lost trying out new Xcode features or researching VS Code extensions that in practice make me 0.05% more productive. Personally — and I do emphasize this is a personal decision — I‘d rather write a well-spec'd plan and go do something else for 15 minutes. Claude's Plan Mode is exceptional, and that‘s why so many people fall in love with Claude once they try it.[^2] Many engineers would rather be hands-on and guide the software development process step by step — which is very understandable when you consider what flow state looks like. Claude will ask you a ton of questions and interrupt itself more frequently to make sure it's not heading in the wrong direction. That really makes you feel like you're doing heads-down engineering work. So Should I Use Codex or Claude? I think back to coworkers I've had over the years, and their varying preferences. Some people couldn't start coding until they had a checklist of everything they needed to do to solve a problem. Others would dive right in and prototype to learn about the space they would be operating in. The tools we use to build are moving fast and can be hard to keep up with, but we've been blessed with a plethora of choices. The good news is that there is no wrong choice when it comes to AI. That's why I don't dismiss people who live in Claude Code, even though I personally prefer Codex. The tool you choose should match how you work, not the other way around. If you use Claude, I'd suggest trying Codex for a week to see if maybe you're a Codex person and didn't know it. And if you use Codex, I'd recommend trying Claude Code for a week to see if maybe you're more of a Claude person than you thought. Maybe you'll discover your current approach isn't the best fit for you. Maybe you won't. But I'm confident you'll find that every AI tool has its strengths and weaknesses, and the only way to discover what they are is by using them. [^1]: As you may know I'm a Swift guy, with a sprinkling of TypeScript for anything that touches the web. [^2]: I will argue that Claude's Plan Mode is basically what working with Codex is like all the time, and one of the main reasons I prefer using Codex.43:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqlthz2d",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2025/12/22/codex-vs-claude-code-today/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Programming","Engineering"],"$type":"site.standard.document","title":"Codex vs. Claude Code (Today)","description":"Every AI tool has its strengths and weaknesses, and the only way to discover what they are is by using them.","publishedAt":"2025-12-22T12:00:00.000Z","textContent":"$5e","canonicalUrl":"https://build.ms/2025/12/22/codex-vs-claude-code-today/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqlthz2d","sort_date":"2025-12-22T12:00:00+00:00"}}] 5f:T223c,Did you know Claude Code has a completely customizable status line? !An image of Claude Code with a custom status line I was browsing Anthropic's documentation when I stumbled on a genuinely useful feature. At first I thought the status line would just display some preconfigured information, but it turns out the status line is fully customizable — you can make it show whatever you want. When I discovered that, I knew exactly what I wanted to build. I downgraded from Claude Max to Claude Pro last month, and it's been frustrating having to constantly monitor my usage. I've developed a tic — typing /usage as often as I used to press Ctrl + S in Microsoft Word. What if I could just see my usage stats at a glance? Building Your Status Line In One Prompt Well, that's exactly what I built. Or more accurately, what Claude built for me. It only took about 10 minutes, and here's the prompt you can use to build it even faster: > I would like to build a status line that shows the usage quotas for my current and weekly session limits. This blog post (build.ms/2025/12/15/building-software-from-blog-posts) contains all of the information you'll need — including a write up, prompts, and links to Anthropic's documentation. More importantly, it contains a script you can use as your foundation. And that's it. One prompt, and you're done. Want to customize it further? Just tell Claude how. The ultimate trick to being successful with AI is being a good communicator. There's an even bigger lesson in this prompt: If there's something you want Claude, Codex, or Gemini to do — you can point them to blog posts or documentation and let them do the research. Have AI do the research for you, and then it will know how to do the work for you. Want to see how I built this from scratch? Keep reading. (And if you're a hands-on learner who wants to level up your AI skills by building your own ideas, this is exactly the kind of technique I teach in my AI workshops.) Building This From Scratch I'll be honest — I didn't write any of this code myself. I just asked Claude to do it. Here's the prompt I started with: > Can you update our status line to display the total context used and how much of our session/weekly quota we've used? Here is some documentation for configuring your status line. https://code.claude.com/docs/en/statusline And it worked! Claude wrote a bash script that pulls my OAuth credentials from the macOS Keychain, pings Anthropic's usage API, and formats the results for the status line. I looked over the code and could immediately see that it was solid. Then I asked Claude to document everything — the script, how it works, installation steps, troubleshooting tips, and anything else that was relevant. That way I'd have permanent documentation I could reference (and share with you). That was as simple as asking Claude: > Let's assume I'm sharing how to do this with a coworker. Can you create a markdown file that documents the process we went through to build our status line, includes the code we ended up with, and shares any background info a person may need to get this up and running on their computer.

Expand to see Claude's technical writeup

bash #!/bin/bash Status line script for Claude Code Displays: Session quota, Weekly quota With comprehensive error handling for authentication issues Read JSON input from stdin input=$(cat) Initialize variables session_pct="" weekly_pct="" Fetch credentials from macOS Keychain CREDS=$(security find-generic-password -s "Claude Code-credentials" -w 2>/dev/null) Validate credentials exist if [ -z "$CREDS" ] || [ "$CREDS" = "null" ]; then echo "OAuth token not found — run /login to authenticate" exit 0 fi Extract access token from JSON ACCESS_TOKEN=$(echo "$CREDS" | jq -r '.claudeAiOauth.accessToken' 2>/dev/null) Validate token extraction succeeded if [ -z "$ACCESS_TOKEN" ] || [ "$ACCESS_TOKEN" = "null" ]; then echo "Invalid token format — run /login to re-authenticate" exit 0 fi Pre-check: Verify token has required OAuth scope SCOPES=$(echo "$CREDS" | jq -r '.claudeAiOauth.scopes[]' 2>/dev/null) if [[ ! "$SCOPES" =~ "user:profile" ]]; then echo "Token missing required user:profile scope — run /login to update permissions" exit 0 fi Fetch usage data from Anthropic OAuth API Note: We omit Accept-Encoding to avoid compression issues with curl USAGE_DATA=$(curl -s "https://api.anthropic.com/api/oauth/usage" \ -H "Accept: application/json, text/plain, /" \ -H "Authorization: Bearer $ACCESS_TOKEN" \ -H "anthropic-beta: oauth-2025-04-20" 2>/dev/null) Validate API response exists if [ -z "$USAGE_DATA" ]; then echo "Failed to fetch usage data — check connection" exit 0 fi Check for API errors in response error_type=$(echo "$USAGE_DATA" | jq -r '.error.type // empty' 2>/dev/null) if [ -n "$error_type" ]; then Handle specific error types if [ "$error_type" = "permission_error" ]; then error_msg=$(echo "$USAGE_DATA" | jq -r '.error.message // empty' 2>/dev/null) if [[ "$error_msg" == "user:profile" ]]; then echo "Token missing user:profile scope — run /login" exit 0 fi echo "Permission error — run /login to re-authenticate" exit 0 else echo "API error: $error_type — run /login if issue persists" exit 0 fi fi Extract usage percentages session_pct=$(echo "$USAGE_DATA" | jq -r '.five_hour.utilization // empty' 2>/dev/null) weekly_pct=$(echo "$USAGE_DATA" | jq -r '.seven_day.utilization // empty' 2>/dev/null) Check if seven_day field is actually null (not just 0) seven_day_raw=$(echo "$USAGE_DATA" | jq -r '.seven_day' 2>/dev/null) Helper function to format percentage format_percentage() { local pct=$1 if [ -z "$pct" ] || [ "$pct" = "null" ]; then echo "0%" else printf "%.0f%%" "$pct" 2>/dev/null || echo "${pct}%" fi } Format the output components session_display="Session: $(format_percentage $session_pct)" Handle weekly display - only show if weekly tracking is enabled if [ "$seven_day_raw" = "null" ]; then No weekly tracking - show session only echo "${session_display}" else Weekly tracking enabled - show both weekly_display="Weekly: $(format_percentage $weekly_pct)" echo "${session_display} | ${weekly_display}" fi bash chmod +x ~/.claude/statusline-command.sh json { "statusLine": { "type": "command", "command": "/Users/YOUR_USERNAME/.claude/statusline-command.sh" } } bash echo '{}' | ~/.claude/statusline-command.sh bash CREDS=$(security find-generic-password -s "Claude Code-credentials" -w) echo "$CREDS" | jq -r '.claudeAiOauth.scopes' json ["user:inference", "user:profile"] /logout /login Session: 39% | Weekly: 27% https://api.anthropic.com/api/oauth/usage json { "five_hour": { "utilization": 39.0, "resets_at": "2025-12-12T20:59:59.707736+00:00" }, "seven_day": { "utilization": 27.0, "resets_at": "2025-12-16T03:59:59.707754+00:00" } } bash CREDS=$(security find-generic-password -s "Claude Code-credentials" -w) ACCESS_TOKEN=$(echo "$CREDS" | jq -r '.claudeAiOauth.accessToken') curl -s "https://api.anthropic.com/api/oauth/usage" \ -H "Accept: application/json, text/plain, /" \ -H "Authorization: Bearer $ACCESS_TOKEN" \ -H "anthropic-beta: oauth-2025-04-20" | jq '.seven_day' bash security find-generic-password -s "Claude Code-credentials" -w bash CREDS=$(security find-generic-password -s "Claude Code-credentials" -w) ACCESS_TOKEN=$(echo "$CREDS" | jq -r '.claudeAiOauth.accessToken') curl -s "https://api.anthropic.com/api/oauth/usage" \ -H "Accept: application/json, text/plain, /" \ -H "Authorization: Bearer $ACCESS_TOKEN" \ -H "anthropic-beta: oauth-2025-04-20" | jq . bash Replace this macOS-specific line: CREDS=$(security find-generic-password -s "Claude Code-credentials" -w 2>/dev/null) With this Linux-compatible line: CREDS=$(cat ~/.claude/.credentials.json 2>/dev/null)

--- Let's take a minute to talk through what happened here: I had a problem, I asked Claude to solve it, and then I asked Claude to document the solution so others could solve it too. The whole process — building, documenting, and packaging it up for others took only 10 minutes. That's the power of AI-assisted development. Using tools this way is the future of knowledge work, which is why I teach this every chance I get. As I've talked about many times, you don't need to be a programmer to build useful tools anymore. All you have to do is be a good communicator — and just ask.44:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqoobi2i",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2025/12/15/building-software-from-blog-posts/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Tutorial","Engineering"],"$type":"site.standard.document","title":"Building Software From Blog Posts","description":"Did you know you can just point AI at a blog post and have it build a piece of software? That‘s where we are now.","publishedAt":"2025-12-15T12:00:00.000Z","textContent":"$5f","canonicalUrl":"https://build.ms/2025/12/15/building-software-from-blog-posts/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqoobi2i","sort_date":"2025-12-15T12:00:00+00:00"}}] 60:T1339,I once had a girlfriend who loved the game Scribblenauts, and I loved watching her play it on her Nintendo DS. The premise seems simple: the main character Maxwell navigates through levels by solving puzzles. But there's a twist. Need to fly? Type "jetpack" and a jetpack appears. Need to write something? Ask for a pencil. The game has an extensive vocabulary and rewards creativity, so rather than typing "fire", you might as well type "dragon" or "Cthulhu" to summon a fully formed monster with all its physical properties at your disposal. I've mentioned the concept of on-demand software a handful of times, and expanded on it in my recent talk at NSSpain. For that talk, I wanted slides that looked identical to ChatGPT's interface but with my talk's color scheme. I built exactly that in a Claude Artifact, and it took less than five minutes using these prompts. I didn't write any code — I just described my problem and told Claude what I needed. That's what I mean when I say Scribblenauts for Software. Building Your First On-Demand Software Is this chat bubble generator my life's work? Of course not. It's just a simple tool — but humans are tool builders. For a motivated developer, building the tools you need on-demand with the help of AI is empowering. Rather than sitting in Keynote all day manipulating shapes until they look just right, I automated the process by creating a piece of software. Three prompts and five minutes later, I had exactly what I needed. No code, no thinking required — just a usable and reusable piece of software. Don't like how it came out? Ask for something different. Don't like it at all? Start over with a completely different direction, this time telling the AI exactly what you don't want. This XKCD comic about automation used to be funny because it was true. Spending hours to automate a five-minute task was absurd, but the timetable has changed. Now you can spend five minutes building something that would have previously taken a day, or an hour building something that would have taken a week. Not every problem is this simple to describe, but sometimes it really is this easy. Today the cost of making throwaway software that helps you accomplish a specific task is practically zero. Building More Complex On-Demand Software I've spent the last few years bouncing between curl and tools like Paw or Postman to test Plinky's API. I use curl when I need something quick and dirty, and a GUI tool when I want to save and reuse an endpoint. But what if I had a personalized tool made just for Plinky? What if I could simply type Create tag build.ms and have AI create and execute the correct curl request for me? This is Scribblenauts for Software thinking. - I open up OpenAI's Codex and ask it to read through my code for Plinky's API layer to understand all of the endpoints I have. - I ask Codex to generate an OpenAPI spec from my code, so now I have a formal document that describes all my endpoints and parameters. - Codex builds a Python script called plinky-api, using GPT-5 to analyze my command line input and read through the OpenAPI spec to determine which route I'm most likely trying to call. Now when I run plinky-api "create tag build.ms", Codex: - Chooses the right endpoint - Constructs a curl request to add a new tag named build.ms - Executes the curl request for me I now have a personalized tool that tests my API. No fiddling with curl requests, finicky GUI tools, or keeping all of Plinky's API routes in my head. I came up with and built this in 15 minutes, and most of that time I was off doing something else while Codex wrote the code. It took a couple of iterations to perfect, but that's nothing compared to the time I've saved since — let alone making my life a little less annoying. Scribblenauts for Software We truly are at the beginning of an era of personalized on-demand software. Today, you can solve practically any problem by creating your own bespoke software. If you have a ChatGPT subscription, Codex is included for free, and Claude Code is included in Claude subscriptions. This means that right now, you can build any tool you want. This is Scribblenauts for Software, and it's a big deal. Today I'm using AI to build my own tools. Tomorrow developers will use AI to build games with interactions that feel dynamic and realtime.[^1] Give it a decade and many entertainment experiences will be intermediated with AI to feel alive, even when they're artificial in nature. When you can infuse everything with AI, suddenly everything becomes a canvas drawn from unlimited objects — much like Scribblenauts. P.S. Want to learn how to build tools like this? I teach hands-on workshops that guide you through every step of the way. [^1]: I recommend this read about how a brilliant hacker used AI to replace the dialogue in Animal Crossing for GameCube with current events.45:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqrimi2o",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2025/12/1/scribblenauts-for-software/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering"],"$type":"site.standard.document","title":"Scribblenauts For Software","description":"AI lets people build any tool in minutes at practically zero cost. The result? An era of personalized, on-demand software.","publishedAt":"2025-12-01T12:00:00.000Z","textContent":"$60","canonicalUrl":"https://build.ms/2025/12/1/scribblenauts-for-software/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqrimi2o","sort_date":"2025-12-01T12:00:00+00:00"}}] 61:T2734,The short version of this post is that: - I'm publicly launching the AI Workshops I've been running since April. - I've taught dozens of people (technical and non-technical) who swear by the results. - It would mean a lot if you'd check them out — and even better, sign up! (P.S. They're a fantastic use of your company's Learning & Development budget.) But I also want to tell you how this all started, and why I believe I can help you learn something life-changing. My First Workshop On a crisp spring evening last April, my friend Jeremy came by my house after I'd made him a promise. He'd been telling me how AI was changing his life. He'd gotten deep into organizing everything he cared about across a myriad of Claude Projects. He had projects for his health, his finances, and so much more. Each one had complex instructions, contained contextually-relevant documents, and thoughtful prompts. He was clearly motivated to get useful results from AI. So I asked him what I thought was a simple question: "Are you using Claude to code — even a little bit?" He hadn't. Jeremy is a PM who didn't know how to code. But having worked closely with Jeremy at Twitter, I knew he was deeply technical. He understood the product and software development process better than most people, many software developers included. I told him we could treat code as just a detail, and that I'd teach him how to use AI for coding. I wanted Jeremy to feel how powerful it is to build your own ideas, and I'm happy to say it worked. To quote Jeremy a few months later: > Meant to mention that coding (via AI) has become my default activity apart from family time. My now wife has to tell me "don't code today". I’m so grateful I’ve found this new hobby. That night turned out to be my first AI workshop. It Starts With An Idea I told Jeremy to bring an idea — one that he wanted to exist — and the two of us would build it together. I knew diving in was the best way forward. I could spend hours answering Jeremy's questions, but most of the answers would materialize in real time if we just made something he cared about. Jeremy described a meticulous process he has when eating out. He's trying to eat as healthy as possible, so he: - Opens a Claude Project with his health goals, dietary restrictions, most recent bloodwork, and other relevant tests. - Drops in a PDF of the restaurant's menu (or takes photos if there's no PDF) and adds it to a new chat. - Asks which menu items he can eat and which ones to avoid for his health needs. Jeremy wanted to automate this process. I thought the idea was interesting, and figured it would lend itself well to the two of us building a prototype. But he wanted to take it a step further than his current workflow. Instead of dropping a menu into a Claude Project, Jeremy wanted to look up a restaurant like he would with Google Maps and have this health check happen automatically. Now this is thinking like a developer. From Idea to App in 10 Steps 1. We started by setting up the Google Places API, so Jeremy could find restaurants by name. 2. I asked Jeremy to describe his idea in a Google Doc. 3. He wrote a couple of paragraphs with varying levels of detail — basically riffing. 4. I asked him, "if you handed this to a software developer, do you think they'd be able to build this app?" 5. He said no, and I agreed. 6. I showed him how to improve our prompt, and shared examples of great prompts that could serve as inspiration. 7. We began adding detail after detail after detail, until it was a full-blown PRD. Good thing Jeremy is a terrific product manager! 8. Once we had a spec, we let AI loose on the problem. 9. I still remember Jeremy's face as AI began building his app. He was awestruck, watching his idea come to life right in front of him. 10. Jeremy spent the next couple of hours asking me whatever questions came to mind as we made progress. I answered them so I could help him learn — and keep him excited about this mind-opening experience. Trust the Process (And Why It Works) That's still how I teach my workshops. I've since run dozens of them, which has allowed me to refine my technique. I love seeing a person's face light up when they have a lightbulb moment, and my real goal is to have three or four of those moments every session. That's when a person realizes they were this close to productively building software with AI this whole time. All you need to know is what you don't know — and that's where I help best. There's some technique involved, and some technical know-how, but I've found that what people need most is to learn the mental models for working with AI. *I fundamentally believe people are not as far from using AI effectively as they think. You may only be a few questions away from learning a life-changing skill, but figuring out what those questions are is the hard part. If you ask Jeremy though, the reward is even higher than you can imagine when you don't know what you're missing out on. That may be true for you too. AI Workshops with @mergesort You'll learn how to write prompts that actually work, turn ideas into apps, and most importantly — build the confidence to work effectively with AI on your own. What I care most about in this world is helping people become the best version of themselves — however they envision that to be. That's why I personalize every workshop to the attendee or team. I know that real learning only happens when you meet someone where they are, rather than giving them slides to read or videos to watch and hoping they figure out what matters. I've always been a good teacher, but what I've discovered is that I'm quite good at doing this — teaching people a lot and inspiring them to do more in a short period of time. If you don't believe me, you can believe the people who have taken my workshops so far. !Jeremy Browning "Joe's a rare combination of deep technical expertise and aptitude for teaching. Having taught me patiently and cogently about Al development, he's cleared a dam for my personal ideas, which l've begun to bring to life. This workshop is a must for anyone with entrepreneurial desire and the interest to learn from one of the greats in Product Engineering today." !"As someone new to software development, Joe was an incredible teacher. He explained things clearly and always pointed me in the right direction. The session was practical and applicable, but not strictly instructional, encouraging me to think critically about how to apply what I was learning. I left feeling confident that I could use those lessons beyond the workshop, and I'm excited to keep growing and share what I've learned with others in my organization." !"In just a few hours Joe showed me how to map my years of software engineering experience into a modern Al-focused workflow. Together we planned and then built a simple app to record Wordle game results, and along the way I had several "aha!" moments that formed a solid understanding of how to work with these tools. I left feeling confident that I will be able to adapt to whatever new models and features arrive next." !"Joe helped me review my Al workflow and quickly spotted improvements that made an immediate impact. Watching his process as a fellow iOS developer gave me a clear picture of how top engineers use Al to code more effectively. He is an exceptional teacher who asks the right questions, shares valuable insights, and makes complex ideas easy to apply. The workshop left me more confident, efficient, and inspired to push my projects further." I've managed to run all of these workshops before launching publicly because word of mouth spread me so thin that I haven't had time to launch publicly! I spend all my time teaching AI now — whether that's running these workshops, helping non-technical people at a local non-profit (who are building real products despite no previous software experience), or on stage at conferences like NSSpain. My life for the last six months has been teaching AI (and building Plinky with the help of AI), and I'm only getting better at it. Six Months of Teaching (And Building) I'm not the only one building with AI. - Jeremy is building apps for himself, his family, and friends. He recently sent me one of his new apps, and I continue to be proud of what he's done and how far he's progressed. - My workshop attendees often write back within 24-48 hours to tell me they've built something they've been wanting for years — something they would never have made without AI doing the heavy work. (One attendee built an app his wife had wanted for four years* in just a few hours — so I'm available for marriage counseling too.) - A team of three that I mentored at the non-profit have started a small B2B business despite having zero prior coding experience. They've built a real product that solves a real business pain point, and just made their first sale to an enterprise customer. Whether you're a designer, product manager, founder, or just someone with ideas — if you can describe what you want, you can build it. And the workshops work great for developers too; I love going deep and technical on how to work better with AI. Now Let's Talk About You AI has the potential to be the most life-changing invention of our lifetime. The question I want you to ask yourself is: "How will it change your life?" AI is becoming a foundational skill that everyone will need — like literacy or math. But we're only at the beginning of this transition, and many people still struggle to get good results. They often believe AI simply can't do certain things that I know it can, and I can show you how. This is what I teach best: helping curious and creative people solve real problems with AI. I go beyond techniques to teach fundamentals and build your confidence to use AI every day in ways that matter. These skills scale from small tasks to large projects and work across every tool — from Cursor to Claude to46:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqty7723",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2025/10/30/ai-workshops-launch/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Engineering","Design","Product"],"$type":"site.standard.document","title":"The Past, Present, and Future of AI Workshops with @mergesort","description":"Six months teaching people to build with AI taught me one thing: most people are just a few questions away from learning so much.","publishedAt":"2025-10-30T12:00:00.000Z","textContent":"$61","canonicalUrl":"https://build.ms/2025/10/30/ai-workshops-launch/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqty7723","sort_date":"2025-10-30T12:00:00+00:00"}}] 62:T2783,When I tell people how great ChatGPT is (or substitute your favorite AI tool like Claude), the question I hear most often is: “But how do you actually use it?” It’s a great question because, as I explained in this blog post, working with large language models (LLMs) isn’t always intuitive and takes some practice to master. Rather than scrambling to come up with examples on the spot, I decided to document how I used ChatGPT over a four day weekend. While this list isn’t exhaustive or representative of every weekend, it’s a great starting point for anyone curious about what ChatGPT can do. You don’t have to read through every single use case, you may want to skim the post and dive into the ones that strike your fancy. If you aren’t a developer, by all means, skip the coding section. If you don’t need health advice, then scroll past how I’ve used ChatGPT to diagnose my medical issues. There is a wide array of use cases discussed, and I’m sure at least one of these AI-assisted workflows can improve your life. Coding Solving multiple complex coding challenges You’ll see the full results in Boutique 3.0, but I was thrilled with how ChatGPT’s o1 model helped me brainstorm, prototype, and build a feature inspired by Swift Data’s Relationships. This feature has been on my wish list for over a year, but I kept putting it off because of the complexity. Using ChatGPT as a pairing partner, I developed about 10 potential APIs for this feature in under an hour, and ultimately found one that I loved. ChatGPT then built a first version of the implementation, which got me 90% of the way there, requiring only minimal tweaks to fully align with my goals and preferences. I also turned to ChatGPT as a teaching aid to review and understand this pull request. This was a fix for a long-standing bug in Boutique, but I had a bit of a mental block when trying to understand how it resolved the issue. I asked ChatGPT to double-check my assumptions and explain what I was missing. Being a software developer is as much about understanding code as writing it, and having a tool to improve my comprehension has been invaluable. Building a simple app I've been dealing with a prolonged bout of tinnitus, and the only thing that seems to help is listening to specific nature sounds — like a mix of fire crackling and light rain. Unfortunately, the app I was using for this was buggy and frustrating to navigate. I realized I could build a better app in just 20-30 minutes with ChatGPT. So, I extracted the sounds from the app’s IPA file and let ChatGPT help me create a new, more functional app with a better interface. Did I absolutely need to build this? No. But it’s made my life a little easier during a tough time. > Sidenote: This is a perfect example of how software development is changing. The marginal cost of creating software has always been high, but with LLMs and the tools being built around them, it’s only getting lower. Developers should take note. Making bespoke apps on demand is already possible — and this kind of capability will continue to shape the future of the industry. Writing SQL queries for me I often need to analyze how people are using my app, Plinky, to better understand how I can build a product that truly helps them. To make informed, data-driven decisions, I needed to determine the ideal number of links users can save before encountering a paywall. To find the right number, I decided to look at the average and median number of links users save in their first day, first week, and first month. Writing queries for this type of analysis can be tedious, especially with complex database schemas. By providing ChatGPT with my database’s DDL (Data Definition Language), I gave it an accurate understanding of my database’s structure — without sharing any actual data. With that context, ChatGPT generated a useful SQL query in seconds, providing the insights I needed. What normally involves googling SQL syntax and some trial and error was effortless with ChatGPT. Automating tedious tasks Fixing annoyances with one-off Python scripts There are small tasks that annoy the heck out of me, but these days, I often ask ChatGPT to write a script to do them for me. The last script I generated was a command line tool that allows me to pass in a list of simulators I would like to keep available in Xcode, and then have simctl delete all the other simulators for me. This saved me time and hassle. Instead of scrolling through an endless list of unused simulators in Xcode (which get re-added after every update), I was able to quickly fix the problem without spending much time figuring out how to write the script myself. ChatGPT’s ability to generate functional code to solve tedious problems is something I’ll never get enough of. Transforming colors in a color palette While considering new color schemes for Plinky, I wanted to quickly prototype colors without having to spend too much time building out a full color palette. Below is the prompt I used to make this work. > I have four different shades of purple colors. > - Light Mode Normal Contrast: #8E44AD > - Dark Mode Normal Contrast: #9B59B6 > - Light Mode High Contrast: #AD68C9 > - Dark Mode High Contrast: #783A91 > I'm trying to replace this color in my color palette with a new shade of blue. The color is #8ECAE6, and I would like to use that for the Light Mode Normal Contrast option. Can you create the other three colors applying the same transformations that we used for our shades of purple? ChatGPT delivered exactly what I needed: four perfectly tinted shades of blue — #8ECAE6, #9BDFEF, #ADEEFF, #78C0CA — each with slight offsets applied for the respective color modes. It was much faster than if I had attempted to create a whole palette myself, which is important because now I was able to try out a bunch of color schemes and see if any of them fit my needs. Double-checking my work Verifying work I’ve done As a small business owner, I file quarterly taxes, well, every quarter. It’s a tedious task, and the cost of making a mistake is high. Nowadays, after copying data between spreadsheets as I usually do, I paste both the original and updated spreadsheets into ChatGPT and ask it to compare them. While I can’t guarantee ChatGPT catches everything, this highlights any differences or potential errors I might have missed. It has found mistakes before, and while I can’t guarantee it will always help, it’s still better than nothing. This quick check has become an important part of my workflow, giving me a little boost of confidence that I’m getting things right. Writing better copy I like to think I’m a pretty good writer, but I often struggle with microcopy. My penchant for illustrative language doesn’t translate very well to the concise, clear text needed in my apps — especially when it’s crucial to communicate information to users quickly. This is where ChatGPT has been a huge help. I’ll upload a screenshot of a screen I’ve built with my best attempt at writing good copy, and ask how I can make the copy more concise and direct. It was particularly useful during my recent redesign of Plinky’s onboarding messages. With ChatGPT’s suggestions, I’ve been able to fine-tune the language across my app and deliver a better experience for users. Research Music Discovery I have a pretty eclectic musical palate, and I’m always looking for new music to listen to. Apple Music and Spotify give you access to every artist, album, and song imaginable, but when I find something I like, what I want is more of that. Unfortunately, their discovery features aren’t great — you either browse endlessly or let an algorithm recommend music until something catches your ear. This weekend, I was listening to Chad Hugo, and all I could think was, “Damn, this slaps.” I ended up having a whole conversation with ChatGPT about him, asking about who he’s worked with, his musical influences, and for more interesting things I should know about him. Then I asked: “Name five other artists I should check out if I like Chad Hugo.” ChatGPT suggested Pharrell, Daft Punk, Timbaland, The Chemical Brothers, and Mark Ronson. I’ve listened to a bunch of music by three of those artists, and agreed they were spot-on for my taste. That gave me confidence to check out the other two, and led me to discover a bunch of new music I’m now excited to listen to. Identifying a bird on my fire escape My lil cat loves watching birds, and while she’s not the most successful hunter (the window hurts), she remains fascinated and convinced she’ll get them one day. I noticed a bird on my fire escape that definitely wasn't a pigeon, and was curious about what bird had such a beautiful speckled chest. With just one photo ChatGPT was able to correctly identify the bird as a European Starling, which it turns out is a lot more common in New York City than I expected. Note: I did a Google Image search to confirm that the bird was indeed a European Starling, but wouldn’t have known where to start if I hadn’t asked ChatGPT in the first place. Answering menu questions I was out at a restaurant, and the menu contained a handful of items I was unfamiliar with. I wanted to find out what they were and if they sounded good, so while I could have googled each item individually, instead I took a picture of the menu and asked ChatGPT about three dishes at once. Asking about the causes of poor coordination I was walking down the street when I had a random thought: what causes some people to have poor coordination? I decided to ask ChatGPT: > What causes a person to have poor coordination? Is it usually neurological, biomechanical, or something else entirely?” The response was very insightful and sparked even more questions. ChatGPT explained various causes, including neurological issues, muscle imbalances, and even inner-ear problems. Each answer led to new questions, and before I knew it, I’d spent 10 minutes learning how 47:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqyvhf2a",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2025/1/1/how-i-used-chatgpt/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Productivity"],"$type":"site.standard.document","title":"How I Used ChatGPT Over A Long Weekend","description":"Real-world examples of how I used ChatGPT over a four-day weekend, providing practical insights for anyone curious about working with large language models.","publishedAt":"2025-01-01T12:00:00.000Z","textContent":"$62","canonicalUrl":"https://build.ms/2025/1/1/how-i-used-chatgpt/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaqyvhf2a","sort_date":"2025-01-01T12:00:00+00:00"}}] 63:T1665,When it comes to creative work, personal growth is rarely a straight path. Whether you’re a developer, designer, writer, or artist, the question of what to focus on can feel daunting. Over the years, I’ve developed a personal framework to help me stay on track: The Rule of Small and Large. The Rule of Small and Large is simple: Every project I take on is either small or large — I avoid medium-sized projects at all costs. Why? Small projects let you learn quickly, experiment freely, and explore new ideas with minimal risk. Large projects, on the other hand, challenge you to go deep, tackle complex problems, and build something truly meaningful. Medium projects often fall into a frustrating trap — they demand too much effort for too little reward. Small Small projects allow you to experiment quickly, see results, and learn with minimal effort. They’re perfect for testing ideas in a low-risk environment. For example, if I want to explore how a programming language solves a particular problem, I might build a quick prototype. The simpler the process, the more likely I am to start. If I had to download an IDE, compile the language from scratch, and read lengthy documentation, I’d probably abandon the idea altogether. The beauty of small projects is that they yield small but cumulative rewards. Each one builds on the last, expanding your skillset and perspective over time. Small projects also thrive on simplicity — lowering the barrier to entry makes it easier to start, which encourages creativity and experimentation. This principle doesn’t just help beginners, even experienced creators benefit from the freedom to explore without overcommitting. By keeping the stakes low, small projects let you test ideas, fail fast, and stay in a constant state of discovery. As an indie developer, I rely on small projects to stay self-sufficient. Whether I’m testing out a new design trend, experimenting with a marketing framework, or playing around with some unfamiliar code, these low-effort experiments validate ideas and help me grow. Small projects aren’t just about learning — they’re about exploring new territory and fueling the creative process. Large Large projects are for those big, ambitious ideas that go far beyond a quick prototype. They start with a vision — a real problem you’re determined to solve. But large projects are rarely straightforward. They demand that you solve countless smaller problems before you even get close to your end goal. It’s this process of breaking down complexity and overcoming challenges that makes large projects so rewarding. The deeper you go, the more you learn. Solving real problems introduces you to complexities you would never encounter in a small project. Let’s say you’re building the next big app for restaurant bookings. At a minimum, you’ll need to understand how time works. That might sound simple, but time zones aren’t so straightforward. Nor are addresses, prices, or payments. Large projects immerse you in real-world challenges that don't fit neatly in a small project, sharpening your skills and teaching you lessons you wouldn’t gain otherwise. That’s the magic of large projects: they push you to go deep, master complexity, and create something valuable. If you have a goal, passion, and are willing to push your boundaries, the impact of a large project can be transformative. Medium If small projects provide you with breadth and large projects provide you with depth, what do medium-sized projects provide you? A trap. Medium-sized projects sit in an uncomfortable middle ground. They require more effort than small projects but don’t offer the depth or rewards of large projects. You might learn something, but it’s rarely enough to justify the energy spent. In my experience, medium-sized projects are the ones that most often leave me frustrated — or worse, burned out. This is why passion and purpose are so important. Large projects demand intrinsic motivation to push through their challenges, and if that motivation isn’t there, it’s often a sign your project isn’t worth scaling up. Medium projects tend to feel aimless — they’re not quite big enough to inspire you, but too involved to finish quickly. When you notice a lack of passion, it’s better to keep a project small, where the stakes are lower and the lessons come faster. Avoiding medium projects doesn’t mean abandoning ideas altogether — it means being intentional about the scope of your work. If something doesn’t excite you enough to warrant a large commitment, let it stay small. Forcing yourself to choose between small and large helps sharpen your focus and ensures that your energy will be spent where it matters most. Small and Large There’s little I love more than starting a new project. Each one feels like a space full of opportunity, a place for my creativity to blossom, and something I can put my heart and soul into. But as exciting as starting can be, I’ve learned that it's important to know upfront where the project is going too. Small projects let me experiment and explore, offering quick lessons and opening my eyes to new ideas. Large projects challenge me to commit deeply, be intricate, and build something meaningful. Both provide a clear sense of purpose, which is why I’m happy to keep building projects small and large. The next time you start a project, pause and ask yourself: is this a small project or a large one? If the answer isn’t clear, you might be stepping into the trap of a medium project. By choosing to stay small or go big, you’ll ensure your energy is spent on work that inspires, challenges, and matters to you.48:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqar3dn62u",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2024/12/16/the-rule-of-small-and-large/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","Industry","Productivity","Programming"],"$type":"site.standard.document","title":"The Rule of Small and Large","description":"A personal framework for choosing projects: focus on either small projects for quick learning or large projects for meaningful impact, while avoiding the medium-sized trap.","publishedAt":"2024-12-16T12:00:00.000Z","textContent":"$63","canonicalUrl":"https://build.ms/2024/12/16/the-rule-of-small-and-large/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqar3dn62u","sort_date":"2024-12-16T12:00:00+00:00"}}] 64:T1bce,People are often disappointed when they use a Large Language Model (LLM) like ChatGPT or Claude for the first time. They've heard about this magical tool that helps people write code, edit blog posts, or suggest delicious recipes, but then Google Gemini tells them to add glue to their pizza, and the magic vanishes. There are countless guides out there on how to improve your prompts or get better answers. But tips and tricks only go so far. Instead, I’d like to share an analogy that might help you come up with better prompts using your own judgment. There is no magic in computers, even when using a computer feels magical. Using An LLM LLMs are trained by scouring data from across the internet [^1], and "learning” that information. [^2] Imagine a world map where all this knowledge is stored in different places: financial advice in America, programming knowledge scattered across Asia, writing skills in Ireland, and so on. - You open ChatGPT and ask for help with your portfolio. - ChatGPT doesn’t know whether you mean a writing portfolio or a financial portfolio. - Instead of asking for clarification, it assumes you mean financial advice. It tries to remember that by muttering to itself that this is a financial question, so from now on we're going to look for all of our information in “America”. - Next, you hand it a writing assignment about the stock market, hoping it will improve the quality of your writing. - But instead of focusing on writing quality, ChatGPT invents some information about stocks. - ChatGPT has just hallucinated). Why LLMs Make Mistakes To simplify, LLMs don’t actually know anything. Instead they take best guesses word by word [^3] to try and predict the best next word, until it completes an answer to your question. A hallucination is a model's best guess gone awry, when the model doesn't have a high quality answer or has misinterpreted some context. Instead of pausing to think — because models can't think — the model will instead continue on with its answer. To humans this reads as incorrect information, bullshitting, or even lying. But the model doesn't really know good from bad, it just makes a statistically reasonable guess. [^4] This is why people sometimes think LLMs are useless — and it’s hard to blame them! But I’d like to show you how to make them work better for you, so you can get more value out of them. Polite People Are Prompt Rule #1 of successful prompting is to be as specific as possible. The more hints you give the model about what you want to know and where to look, the better its answers will be. In our earlier example, the model struggled because it couldn’t tell if you meant a financial or a writing portfolio, so it looked in the wrong place. The frustrating part? The model had the information you needed but it didn’t know where to find it. Let’s extend our map analogy a bit to show why hints are useful. Imagine you’re solving a programming problem. As we discussed, programming knowledge is scattered across Asia. The model knows most JavaScript expertise resides in Japan, while Swift knowledge is mostly in Korea — though there are exceptions. When we tell the model we’re looking to solve a JavaScript programming problem, we’re helping it focus on information in Japan, not South Korea. This avoids the earlier mistake of the model searching for writing advice in America, which is good because America ranks 36th in the world for literacy. We can narrow it down further by specifying we want to use d3.js, guiding the model to look in Tokyo — specifically, Shibuya, where d3.js information “lives.” Using An LLM (Better) Let’s say I have another programming problem I want to solve. I give ChatGPT a code snippet that draws a simple multi-colored square on the screen. Then I tell ChatGPT: > I want to make this darker - First of all, what is this? The model has no idea. - What color do you want to make darker? - What does “darker” even mean? Do you want the multi-colored box to be one solid color, like black? Because that's as dark as anything can be. - Follow-up questions: How much darker? Do you have a specific percentage in mind? Which method should it use — adjust hue, saturation, RGB, luminance, etc? If a human would have follow-up questions for you, a model probably will too. You can answer those questions in advance or even instruct the LLM to ask you questions if anything is too ambiguous or needs clarification. --- Now that we’ve anticipated some potential issues, we can compose a clearer prompt to describe the problem we’re trying to solve. > I have a JavaScript code snippet that draws a simple multi-colored square on the screen. The colors are currently red, blue, green, and orange. I want a new color palette that’s slightly darker, but still vibrant so the colors pop. > - Please rewrite the code without using third-party dependencies or frameworks. > - Generate 10 color palette options that I can try in my project. > - Once I review them, I’ll let you know which one I like best, and I’d like 10 more options based on that. > If anything is unclear or if you have any follow up questions that may help resolve ambiguities, please ask them before writing any code. See how much clearer this prompt is? It’s more likely to get a good answer because we were very specific about the problem we’re trying to solve. Even as I wrote it, I realized there were multiple ways to approach this problem and details I hadn’t initially considered. The process is a bit like teaching someone who’s never seen the problem before — talking it through helps clarify what needs to be done, both for them and for you. --- Writing good prompts is like any skill — the more you do it the better you'll get. While this analogy is just a starting point, I hope it demonstrates why good prompting is important and how it can help you make the most of this new, almost magical technology. With this technique we’ve brought a little more determinism into a non-deterministic system. If you’re interested in understanding how LLMs interpret your prompts, you might enjoy this article by Ethan Mollick. For a deeper dive, Stephen Wolfram’s fascinating article provides an in-depth explanation of how LLMs function. Finally, for advanced prompting techniques, this paper is an excellent resource to elevate your skills with LLMs. [^1]: For the purposes of this explanation, let's not get hung up on the fact that this is a legal and ethical gray area. [^2]: For the purposes of this explanation, let's not get hung up on the fact that LLMs are statistical models, and aren't learning anything. [^3]: This in itself is an oversimplification. LLMs actually use tokens, which are a way to chunk letters together, and usually amount to 3-4 characters per token. [^4]: There are a lot of techniques to minimize hallucinations, and to provide better information. This is partially why AI has advanced so much, so quickly, as we've done a great job of preventing LLMs from going off the rails.49:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaraa2y2z",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2024/11/14/prompting-large-lanuage-models/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Productivity"],"$type":"site.standard.document","title":"A Mental Model For Prompting Large Language Models","description":"A practical analogy for understanding and improving how you work with AI tools like ChatGPT and Claude, moving beyond tips and tricks to develop better judgment.","publishedAt":"2024-11-14T12:00:00.000Z","textContent":"$64","canonicalUrl":"https://build.ms/2024/11/14/prompting-large-lanuage-models/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaraa2y2z","sort_date":"2024-11-14T12:00:00+00:00"}}] 65:T13cd,> Note: When I left Twitter to become an indie developer I committed to helping others with my work as much as I could. My goal has always been to make the lives of developers a little easier, so they can turn their ideas into, apps, products, and businesses. Every week this month I will be open sourcing something I've built for Plinky, my very own personal Open Source October. My career rests on the foundation that others have created, and I want to continue paying it forward by lowering the barrier for others to make what matters to them. Now, onto the post. If you came here for the Buttondown Framer Component I've open sourced, you can skip the post below and grab the code on Buttondown’s official repo. But if you happen to be interested in the technical details of how to build a customizable Framer Component, keep reading for a walkthrough. --- When I launched Plinky’s Links You’ll Love newsletter, I knew I needed a beautiful landing page to set the right expectation for quality. The problem? I'm sick and tired of writing HTML, CSS, and JavaScript, with zero interest in building a landing page from scratch. !A frowny penguin I’d heard a lot about Framer from my designer friends, and how easy Framer made it to build beautiful websites. The canvas felt familiar, like a more interactive version of Figma. Within half an hour of playing with Framer, I had a slick landing page with fancy animations that took only minutes to create. Despite a few missing features and minor quirks, I fell in love immediately. Code Components What really sold me on Framer wasn’t just the features it has, but how it handles what it doesn’t have. Framer provides common elements for website layouts, like Stacks, Images, and Text, but sometimes you need something custom. That’s where Code Components come in. At its simplest, a Code Component is a React component, which means you can create just about anything with a bit of JavaScript.[^1] There are also many pre-made components available, and best of all, they're fully customizable. The magic of Framer’s Code Components lies in Property Controls, which are essentially React props. By adding Property Controls, Code Components become as customizable as any other Framer element. I’ll explain how this works in just a moment. !Custom Framer Component Framer x Buttondown I use Buttondown, a popular indie email service, to deliver my newsletter. Buttondown is run by Justin Duke, a prolific writer who's very helpful and shares a lot of valuable lessons — really an all-around great guy. While the default look of Buttondown’s email form is basic, I wanted something more polished for my website. This meant writing a lot of CSS, or… I could create a Code Component. 🤔 I dove into the Framer docs, and once I had an understanding for how they worked, I fed that documentation to Claude to outsource the work of writing a template for my Code Component. The real fun began when I realized I could open source this component for other Framer users, and leaned into my love for building reusable abstractions. Building A Code Component The structure of a Framer Code Component is similar to any other React component. You import React, write functions, and use features like State and Effects. Most of the time you’re just writing standard React so I'll skip over some details, but you can view the complete code for this component here. This component has various customizable properties, like newsletterConfiguration, formStyle, text, making it adaptable to different design needs. But the real power of a Framer Code Component comes from the addPropertyControls function. By adding Property Controls, we can make this form as customizable as any other standard Framer element. This is what allows us to transform the Buttondown form I built for my website, and turn it into a form that anyone can personalize for their website. The addPropertyControls function allows us to define these customization options using Framer primitives like ControlType.String, ControlType.Number, ControlType.Color, and ControlType.Object. These controls turn into beautiful configuration panels in Framer, where users can tweak settings to match their design preferences. !Custom Framer Component Now, we have a fully customizable Framer Component, ready for any Buttondown newsletter, whether your favorite color is red, blue, green, or even my favorite color in Plinky, pink. --- Speaking of newsletters, if you enjoy thought-provoking and enjoyable articles delivered straight to your inbox, I’d love for you to see the form I made for my Buttondown newsletter, and perhaps even sign up.. And if you have any thoughts or suggestions, I'm always available on Threads. I welcome any and all polite discussions. [^1]: Although I started off this post by saying I didn’t want to write JavaScript (or TypeScript), I find writing this kind of code is a lot more enjoyable than laying out a responsive website that works across half a dozen browsers.4a:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqas72mq2o",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2024/10/14/building-a-customizable-framer-component/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","Engineering","Open Source","Tutorial","Web"],"$type":"site.standard.document","title":"Building A Customizable Framer Component","description":"A technical walkthrough of building a customizable Framer Code Component for Buttondown email signups, complete with Property Controls and open source code.","publishedAt":"2024-10-14T12:00:00.000Z","textContent":"$65","canonicalUrl":"https://build.ms/2024/10/14/building-a-customizable-framer-component/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqas72mq2o","sort_date":"2024-10-14T12:00:00+00:00"}}] 66:T26f4,> The post below was written by me, originally featured on the Plinky blog. > To celebrate the launch of Plinky you can get 50% off of a yearly subscription by redeeming this offer: plinky.app/offer/REDPANDA There are few words I've ever said more excitedly than these: I want to tell you about my latest app, Plinky. !Plinky Screenshots Plinky makes it incredibly easy to do something we do every day, save links for later. You may already have a way to save links, I know I've tried every method under the sun, to the point where I decided to build my own app. That app is Plinky, and today it's available to download on the App Store. Over the last 18 months people have been loving Plinky, because it fixes the same problems I ran into when I've tried to save links in the past. - I tried filling up Apple Notes with links to come back to later, but that proved to be a disorganized mess. - I tried leaving browser tabs open with links I wanted to look at later, but the links I needed later got lost amongst the dozens of tabs I had open for right now. - I even tried spending hundreds of hours using apps dedicated for reading articles later, but not all links are for reading. All of that led me to build Plinky, a home for your links. Why I Built An App To Save Your Links I love links. I love reading links, I love it when people send me links to funny videos, I love collecting links, and I love sharing links. Something else I love is my fiancée Colleen. Shortly after meeting Colleen discovered how much I love links, and I'm very grateful that she found that to be an endearing trait. While she loved that I was thinking of her and sharing interesting articles, recipes to cook together, or cute animals I knew she'd like, she quickly found it distracting to have her phone buzzing with all the links I'd share while she was working. She suggested saving the links for later, so we could look at them together at night. That’s when I started working on a small project, a home for these links, an app that would become Plinky. I started to show people the app I'd made for the two of us, and they loved it. They were all saving links they needed but were unhappy with one thing or another. They kept telling me that they wanted an app like this for themselves. A few months later when I left my job at Twitter to start my own company building personal, playful, productivity apps, I decided that the app I needed to build for all of these people was Plinky. Plinky became more than an app to Colleen and I, it became an essential tool in our lives and a ritual. Sometimes Colleen would save a cute picture of a red panda for me, sometimes I would save an activity we could do next weekend, but every night the two of us would come together and share the links that we'd saved for each other. To this day we still have plink time, where we lay in bed and show each other the links we've for each other. --- Links come in all forms. An article just isn't the same as a Tweet or Instagram post you want to show a friend. A YouTube video or TikTok won't even open in an app dedicated to reading. Many apps like LinkedIn or Facebook have their own bookmarking systems, but I've wasted hours bouncing from app to app trying to remember where that link I saw a month ago but need now is. As I’ve built Plinky I've heard people tell me about the myriad of techniques they have for saving and storing links. There are an unending amount of organizational methods people use to reference links they might need later, but none of those perfectly fit into one app. The experience of listening to Plinky's beta testers led me to three realizations: 1. It needs to be easy to save a link, the simpler the better. Saving a link shouldn't take work, it needs to be one tap, so you can easily save a link. 1. People's workflows benefit from having a universal inbox for their links. It should be easy to organize and find links you've saved for later, otherwise why are you even saving them? 3. People aren't always looking for an app dedicated to reading articles. The diversity of links people save means many people are looking for a tool that helps them do anything with the links they've saved, a Swiss army knife for their links. This is where Plinky comes in. Plinky: Saving Links Made Effortless Plinky makes it easy to save a link for later. You can save links from anywhere with just one tap. Whether you’re on the iPhone, iPad, or Mac, in your favorite browser like Chrome, Firefox, and Safari, or want to build personalized workflows using integrations like Zapier, Unread, Shortcuts, or Plinky's API, saving a link for later couldn't be easier. Plinky is highly customizable. You can make the app look, feel, and behave however you like. The links you save are very personal, as is what you want to do with them after they're saved. People have all sorts of workflows, so it’s a core goal of mine to make Plinky easy enough for my mom to use (she really is a big user), yet something you can infinitely tweak to make Plinky a perfect place for your links. Plinky has organizational features like Folders, Tags, Search, and Pinned Links. Each of these provides an opportunity to mold Plinky to your needs, and makes it very easy to find your saved links later. What's Next For Plinky I'm continuing to build upon these three pillars: ease, customization, and organization. Over the coming months Plinky will have the ability to: - Add timed reminders for your links, for the purpose of reviewing a link in a certain time or context. - Import your links from services like GoodLinks, Raindrop, Pocket, and other places you may already have your links saved. - Create Secure Folders, giving you a space for more sensitive links protected by TouchID and FaceID. - Indulge in an elegant reading experience that lets you customize fonts, sizes, and background color for the ideal read it later experience. - Enjoy a native Mac app, one that takes the beautiful design Plinky already provides on iPhone and iPad and makes it feels right at home on the Mac. - Have more ways to build personalized workflows around your links. It's easy to tag a link, but it should always be getting easier. Pinning links helps you prioritize some links over others, but the best experience will be even quicker. Plinky's search is powerful but it can get even more helpful by searching every word of a webpage you save. A Favor, If I May If you've made it this far I want to say thank you to you for caring so much about the app I've made. Plinky is the result of more than 1,500 hours of work. If that doesn't emphasize how much of my heart, sweat, and tears are in Plinky, I'm not sure what will. I'm truly excited about the idea of spending another 1,500 hours making Plinky better for you, then 1,500 more hours after that. If I may ask: - Please download Plinky and try the app. I would love to know what you love about Plinky, what can be improved, or what I can do to make Plinky an integral part of your life. - If you like the app, I would greatly appreciate a rating or review. Ratings decide the success or failure of apps the App Store, and as much as I hate to ask, I would like to succeed because that means I get to continue building an app that strives to make your links better. - Tell your friends. I hope it's not too gauche to ask, but I genuinely believe that Plinky can help improve many people's lives. Every person who tries Plinky may be a person who's now using an app that makes their life a little better, and nothing in this world makes me happier getting a chance to help others. Thank You While you can say that I built this app, nothing in this world happens alone. I've had a lot of help along the way, and want to make sure anyone who's helped even in the smallest way is honored in Plinky. Plinky has a Thank You section on the About Red Panda Club screen where I've assembled everyone who's helped me while I've been building Plinky. !Thank You Screen The rest of this post was written for me and my loved ones, but you’re free to read it as well. I would like to first mention my wonderful fiancée and soon to be wife Colleen. Thank you for supporting me with your love, your advice, your daily feedback, and the hours of hands on help you provided over the last few months in creating and solidifying Plinky's launch plan. I love you, and because of you my dream is now a reality. Thank you to my family for cheering me on this whole time, without you I wouldn't be the person I am today. Plinky is a personal expression of who I am and the way I want to help make the world a slightly easier to navigate place. If it wasn't for you I wouldn't have that perspective, so thank you for instilling those values in me. Thank you to Joey Banks for being a great friend, for our weekly calls, and for your help in shaping Plinky's App Store screenshots and website. Because of you I was able to show the world what I've built in the way that reflects what I wanted people to see. While I'm here bragging about my friend, if you're looking for the world's foremost expert in Figma, you should reach out to Joey. Thank you to all of the beta testers who have provided endless ideas, feedback, and guidance over the last 18 months. Plinky wouldn't look or work the way it does today without your amazing ideas, and of course your extremely helpful bug reports. If you've provided a suggestion during Plinky's beta then odds are it's already been implemented, or is on the Plinky roadmap I've constructed for the next year and beyond. Thank you to anyone who's helped me over the last few years, whether or not that involves Plinky. The last few years haven't been easy, and while I can't always be strong, I can't always be brave, what I can always be is grateful and appreciative. Now if I may leave you with a few words: Be good to others, help those in need, and save a few links.4b:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaslbaq2o",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2024/5/9/introducing-plinky/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","Apps","iOS"],"$type":"site.standard.document","title":"Introducing Plinky: My Love Letter To Links","description":"The story behind Plinky, my personal app for saving and organizing links that makes it incredibly easy to save links for later.","publishedAt":"2024-05-09T12:00:00.000Z","textContent":"$66","canonicalUrl":"https://build.ms/2024/5/9/introducing-plinky/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqaslbaq2o","sort_date":"2024-05-09T12:00:00+00:00"}}] 67:T204e,The Turing test is dead, and we killed it. The Turing test is a test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. From the 1940s [^1] to the 2010s people programmed computers, and computers could only do what they were programmed to do in a rules-based deterministic manner. Sometimes a person would program the computer and it would do something unexpected, but 100 out of 100 times the computer was doing what it was programmed to do whether the person liked it or not. While there has been experimentation with what today we call AI since the 1950s, those machines were a long ways away from passing the Turing test. Why does using ChatGPT feel more like a conversation with the smartest person you know than a computer? It's because ChatGPT doesn't solve problems deterministically the way a programmed computer does, it solves them probabilistically. [^2] ChatGPT demonstrates the ability to think about something in a logical, sensible way, the definition of reasoning. [^3] We've created something completely new here, a reasoning computer. [^4] Working With A Reasoning Computer There are so many political, societal, economic, and ethical implications of Large Language Models (LLMs), 5,000 words wouldn’t be enough to cover all those thoughts. (Trust me, there’s a much longer post sitting in my drafts.) But what’s really captivated me is why a reasoning computer really is different than anything we’ve used before, a conclusion I could only arrive at through experience. ChatGPT has been an essential tool for me over the last month, especially over the last week as I've been building Plinky's browser extension. I'm a very experienced iOS developer but have little experience with web development. I know enough TypeScript and React to cobble together something with lots of help and guidance, but it will take me much longer than someone who knows what they're doing. A browser extension is important for Plinky to be successful though, which presents a unique challenge: I know what I want, I know how to describe it, I don't quite know how to get it, but I will know when ChatGPT gives me the wrong answer so with some nudging I can get what I'm looking for. Here's why the process of pairing with ChatGPT works, and how it helped me build a fully functional browser extension that lives up to my standards in less than a week. (With far less frustration than if you took away the tool and gave me a whole month.) 1. A simple browser extension to save links to Plinky's database is a much smaller problem than building a whole app. The problem is self-contained, which makes it quick and easy to test ChatGPT’s results and see if the output matches my expectations. In fields like mathematics or computer science it's generally easier to verify a solution's correctness than come up with a solution in the first place. 2. I may be a novice web developer but I'm a great programmer. Even in a domain where I’m not comfortable I can describe the problem I'm trying to solve, assess whether a solution is good, do some research (on my own or with the aid of Perplexity and ChatGPT), and nudge the reasoning computer in the right direction. 3. This isn't a process where I ask for something and am given exactly what I want, but I can promise you it's much easier than becoming a good enough TypeScript developer to build the high quality browser extension I want. 4. Little by little the browser extension looks and works more and more how I want it to be, until it does exactly what I want it to do. 5. The whole process is interactive so I’m learning about how to get to the right solution. Not only do I have what I want, but this iteration made me a better web developer, I started off only knowing what the wrong output looks like but now I also know how the correct solution should look. --- This is just one example of how I was able to accomplish something I previously wouldn't have been able to do thanks to an LLM, the number of tasks I turn to LLMs for is growing every day. The same way that GPS becoming ever-present means I haven't opened a map in almost two decades, I find myself turning to ChatGPT or Perplexity rather than opening Google and clicking a bunch of links to find answers. I used to do my own research, I used to be the reasoning machine, but now I'm offloading more and more of that work to Large Language Models. How Can A Reasoning Computer Even Work? People will say that ChatGPT can't do math, and that's true in the most literal sense. A Large Language Model may not know what addition and subtraction mean to a human, but it can use tools and synthesize the correct results to add and subtract numbers better than a person. Similarly people point out that ChatGPT can't read, because it's just a stochastic parrot that means it can't provide intelligible output. It's true that LLMs are complex statistical models, yet despite ChatGPT not knowing English from Urdu the way people do it's still capable of translating from English to Urdu to Russian to French in a way that I never would be able to. The fact that Github Copilot [^5] doesn't actually know the difference between JavaScript and Swift hasn't stopped it from making programmers 55% faster at coding. Large Language Models use a different form of problem solving that starts with inputs and extrapolates technique. That's the reverse of how humans believe they develop their skills, if you study hard, read a lot, and put in enough hours as a writer you too can become the next Faulkner or Shakespeare. But think about the way you first learned your native language, you listened and watched the world around you for 1-2 years, then reverse-engineered how the technique works. We're reasoning machines too, the difference is that the entirety of the internet wasn't preloaded into our brains the way it was into an LLM. (For the best, I don't know if you know but there's some bad shit on the internet.) When we say ChatGPT can't do this or ChatGPT can't do that what we're doing is anthropomorphizing flaws onto the system, derived from our own experiences of solving problems successfully. The problem solving process may be difficult for people to understand because this is the first computer that doesn't do exactly what you tell it to do. Our intuitions may view this as a flaw, but OpenAI loading the whole internet into ChatGPT and creating a simple model for how to think rather than directly programming the machine is the reason this computer is incredibly useful in new and previously unexplored ways. --- Simon Willison says that these tools make you more ambitious with what you can accomplish, and I'd like to build upon his axiom. When you have a reasoning computer you only have to know what the wrong result looks like, not how to get the right result, and that alone has the power to change how society solves problems. [^1]: Ada Lovelace deserves credit for writing the world's first computer program 100 years before ENIAC, but in this context I'm using the timeframe of the 1940s to focus the post on generally programmable computers. [^2]: It's perfectly fair to debate whether this is how the inner-machinations of ChatGPT work, but I feel very strongly that at a minimum you can say this about the output ChatGPT provides. [^3]: This isn’t because ChatGPT is sentient, but in all likelihood because it was trained on a corpus of human-generated data. It's difficult to define "thinking" in this context, my personal view is that there is no thinking without sentience, but in this context what I call thinking isn't the low-level internal machinations of ChatGPT, but one level higher — the step by step token output process that people using ChatGPT see in the process of getting their result. [^4]: I'd like to co-credit Joe Ugowe with coining this term, it stemmed from a wide-reaching discussion we had last night about our experiences with ChatGPT and Large Language Models. [^5]: Github Copilot is a Large Language Model product like ChatGPT, but trained with a coding-specific focus, which allows it to be integrated into a whole suite of Microsoft's programming-related tools and platforms.4c:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqasnovv2a",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2023/11/10/the-reasoning-computer/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","AI","Industry","Programming"],"$type":"site.standard.document","title":"The Reasoning Computer","description":"How will we solve problems in a new era of computation, where the very foundations of how a computer is expected to work are no longer deterministic?","publishedAt":"2023-11-10T12:00:00.000Z","textContent":"$67","canonicalUrl":"https://build.ms/2023/11/10/the-reasoning-computer/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqasnovv2a","sort_date":"2023-11-10T12:00:00+00:00"}}] 68:Te8d,
Is coding a science or an art?
— Matthew Garlington (@dejadu13) May 21, 2022
I've seen many versions of this question posed over the years, and to Matthew's credit it's a very good question. As you can see in the replies people translate their lived experience writing code and answer art or science based on however they conceptualize and practice programming. A few years ago MIT conducted a study that concluded "reading computer code is not the same as reading language", answering the question of whether coding is art or science with a rigorously documented "both". While I'm hard-pressed to argue with science, I'd like to provide a different answer, one that's a little more conceptual. The instinct to qualify coding as art comes from the practice of programming. While writing a program software developers make many choices, ones that rarely have hard and fast rules. As long as your program compiles and runs as expected, you can make any choice you want. It's all human input, if you ask 100 developers to write a complex program they'll write it 100 different ways. Heck, if you ask one developer 100 times they may write it 100 different ways. Code can create something beautiful, enable an amazing experience, and people will even call a piece of code elegant or exquisite. To someone in the arts this all sounds very familiar. The instinct to qualify coding as a science comes from the output of a program. Developers desire deterministic results for their programs, when you provide an input to a program you expect to receive the same output every time. The practice of programming itself can be imprecise, and the same way that science operates in unknown space, building a program can often feel the same. Developers will try to minimize the ambiguity by using industry-tested practices, much the same that a scientist uses a standardized beaker rather than throwing some chemicals in an unclean cup and seeing what happens. When put together that truly sounds like the practice of science. What we've found ourselves asking is whether code is a paint brush or a calculator, and frustratingly the answer appears to be both. Despite the fact that programming looks like art and looks like science, I still think there's a concept that better fits the practice of coding, creative expression. We can look to writing as a point of comparison. Sometimes it's hard to believe that the most beautiful poem uses the same medium as boring technical documentation because the artifacts look and feel so different, but we all know it's true. The medium doesn't impose limitations on how someone can choose to express themselves, it's a tool for creative expression. With writing it's impossible to separate the medium from the artifact created, and the same is true for code. Code can be artistic and create something new in the world, or it can simply exist to accomplish a task. Code helps people capture the meaningful moments of their lives, lets you carry the beauty of every song ever recorded in your pocket, but it's also the boring spreadsheet that helps you run your business. So is coding an art, a science, or creative expression? I say coding isn’t science, it's not art, it’s not quite a craft, but a malleable form of creative expression. When given an infinitely flexible and manipulable canvas people use their imagination to create wondrous things, ones not limited to simple categorization.4d:["$","$L30","at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqat3u6w2s",{"publication":{"href":"https://build.ms","pubRecord":{"url":"https://build.ms","icon":{"ref":{"$link":"bafkreiggonss67wt7o2zqgowfhaspt35us2xlzf3ucdpksxlhdrxjyiioi"},"size":7922,"$type":"blob","mimeType":"image/png"},"name":"Build.ms","$type":"site.standard.publication","basicTheme":{"$type":"site.standard.theme.basic","accent":{"b":228,"g":173,"r":38,"$type":"site.standard.theme.color#rgb"},"background":{"b":231,"g":248,"r":255,"$type":"site.standard.theme.color#rgb"},"foreground":{"b":51,"g":18,"r":0,"$type":"site.standard.theme.color#rgb"},"accentForeground":{"b":255,"g":255,"r":255,"$type":"site.standard.theme.color#rgb"}},"description":"Lessons about engineering, product development, design, business, marketing, and AI.","preferences":{"showComments":true,"showMentions":true,"showPrevNext":false,"showInDiscover":true,"showRecommends":true},"theme":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d"},"ownerDid":"@mergesort.me","contributors":[{"did":"did:plc:b6eke66r3vbmnegg73qgprl6","handle":"mergesort.me","displayName":"Joe Fabisevich"}],"documents":{"comments_on_documents":[{"count":0}],"document_mentions_in_bsky":[{"count":0}],"recommends_on_documents":[{"count":0}],"data":{"path":"/2022/5/21/coding-as-creative-expression/","site":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.publication/3mmypfmg4sx2d","tags":["Top","Engineering","Programming"],"$type":"site.standard.document","title":"Coding As Creative Expression","description":"Exploring whether coding is art or science, and how programming serves as a creative medium for solving problems and expressing ideas through code.","publishedAt":"2022-05-21T12:00:00.000Z","textContent":"$68","canonicalUrl":"https://build.ms/2022/5/21/coding-as-creative-expression/","theme":"$undefined","preferences":"$undefined"},"uri":"at://did:plc:b6eke66r3vbmnegg73qgprl6/site.standard.document/3mmyqat3u6w2s","sort_date":"2022-05-21T12:00:00+00:00"}}]

Tag: TOP

Datensouveränität beginnt bei deiner digitalen Heimat

Your Agent Deserves Logs

Being A 1.5-10x Developer

WordPress 7.0 „Armstrong“ ist da: Warum das beste Update das ist, von dem du kaum etwas merkst.

The Teaching Computer

The Claude Code Leak

Workslop

Small Steps For Agent-Friendly Codebases

The Hidden Cost Of Communication (A Case Against Big Companies)

Automating Codex

Supporting Markdown Search For LLMs

Updating Your Website For Generative Engine Optimization

AI Agents Are Starting To Eat SaaS (Really)

The Democratization of AI

The Claude Code Moment

Build.ms: Making Technology Accessible To Everyone

AI (Without the Hype)

Codex vs. Claude Code (Today)

Building Software From Blog Posts

Scribblenauts For Software

The Past, Present, and Future of AI Workshops with @mergesort

How I Used ChatGPT Over A Long Weekend

The Rule of Small and Large

A Mental Model For Prompting Large Language Models

Building A Customizable Framer Component

Introducing Plinky: My Love Letter To Links

The Reasoning Computer

Coding As Creative Expression