One of the easiest mistakes to make with AI video is asking for too much.
A character walks through a room. Curtains move behind them. The camera pushes forward. Light changes across the wall. Steam rises from a cup. Someone turns toward the window.
Individually, none of these instructions sounds unreasonable.
Together, they can make a ten-second clip feel strangely restless.
I noticed this while experimenting with Seedance 2.5 and thinking about how I write prompts for short video. My first instinct was usually to describe everything I wanted to see. Lately, I have been doing almost the opposite: deciding what doesn't need to move.
Motion needs a hierarchy
Consider a simple shot of someone sitting beside a café window.
There are dozens of possible movements:
the person can look up;
a hand can reach for the cup;
pedestrians can pass outside;
rain can run down the glass;
the camera can slowly move closer;
reflections can shift;
steam can rise from the drink.
Put all of them into one shot and there is no obvious place to look.
Instead, I now try to choose one primary movement.
Maybe the entire purpose of the shot is the hand reaching for the cup.
Everything else becomes secondary. The camera can remain nearly still. Background pedestrians do not need noticeable actions. Even the rain can become something we register rather than something demanding attention.
The scene becomes quieter, but it often feels more intentional.
“Keep still” is a creative decision
This was the part I initially underestimated.
When working with generative video, it is tempting to think primarily in verbs: turn, walk, pan, reveal, float, rotate, zoom.
But stillness carries information too.
A locked camera can make a small character movement more noticeable. A quiet background can make an object entering the frame feel important. Holding a composition for another second can give viewers enough time to understand what they are seeing.
That means prompting is partly an exercise in subtraction.
With a workflow such as Seedance 2.5, I find it more useful to establish the visual idea first and then ask which movement actually advances the shot. Reference images can help define the scene, but they do not make that decision for you.
I use a simple three-level test
Before generating a scene, I separate motion into three levels:
Primary motion — What should the viewer notice first?
Supporting motion — What makes the scene feel alive without stealing attention?
Still elements — What should remain visually stable?
For the café example, that might become:
Primary: the hand slowly reaches for the cup.
Supporting: subtle steam from the drink.
Still: camera, furniture, background composition.
This takes less time to write than a long cinematic prompt, yet it gives me a much clearer idea of what I am actually testing.
If the generation fails, I can also identify the problem more easily.
Not every shot needs to prove what the model can do
There is another reason I like this approach.
AI video demos naturally gravitate toward spectacle. Large camera movements, transformations, particles and complicated action are effective ways to demonstrate technical capability.
But most useful video is not a demo.
A product clip may simply need an object to remain recognizable while the camera moves around it. An editorial sequence may need a few seconds of visual breathing room. A short story may depend more on the pause before an action than the action itself.
The appropriate amount of motion depends on what the shot is supposed to communicate.
A better question before generating
I used to finish a prompt and ask:
What else could happen in this shot?
Now I find another question more useful:
What can I remove without losing the idea?
If removing a camera move makes the action clearer, remove it.
If the background animation adds nothing, keep the background quiet.
If one gesture communicates the entire moment, give that gesture enough time.
Generative video makes it increasingly easy to put motion everywhere.
Learning where not to put it may be the more interesting creative skill.