Fundamentals
How to write an AI video prompt that works
The structure that separates a prompt producing something usable from one producing a pretty, generic clip.
01
Describe the shot, not the subject
The most common mistake is writing a description of a thing rather than a description of a shot. "A luxury watch on a table" tells a model what exists. It says nothing about how it is seen, and the model has to invent the rest — which is why the output looks like stock footage.
A shot description answers four questions: where is the camera, what is it doing, what is lighting the subject, and what happens over time. "Extreme macro on a mechanical watch movement, hard light raking across the surface at a shallow angle, camera drifting slowly left, a single specular highlight travelling along a polished edge" answers all four. Same subject, entirely different result.
02
Give it a beat structure
Video has a dimension photography does not: time. A prompt that describes only a static image gives the model no reason for anything to change, and you get a clip where a camera drifts aimlessly for five seconds.
Write the beats explicitly. "Open on X. On the second beat, Y enters. Hold the final frame still for two seconds." Even when a generator does not follow the timing precisely, a prompt with a described structure produces a clip with structure, and one without produces a clip without.
03
Name the lighting setup
Lighting is the single highest-leverage thing you can specify, and almost nobody does. "Backlit with a large soft source behind the subject and no fill, so the shadow side falls to near black" changes the output more than any amount of adjectives about mood.
If you do not know the vocabulary, describe the physical situation instead: one window on the left, nothing else. Late afternoon sun coming through blinds. A single overhead bulb in an otherwise dark room. Physical descriptions work because they are unambiguous.
04
Say what you do not want, in the positive
Negative instructions are unreliable — many video models do not have a separate negative channel, and "no text" in the prompt body sometimes produces text. Where you can, phrase the constraint positively: "clean seamless backdrop" rather than "no clutter"; "a single subject centred in frame" rather than "no other people".
Where a negative is genuinely necessary, put it at the very end and keep it short. A long list of negatives dilutes everything before it.
05
Change one thing at a time
When a result is close but wrong, the temptation is to rewrite the whole prompt. Resist it. Change one element — the lighting, or the camera move, or the lens — and regenerate. If you change three things and it improves, you have learned nothing about which one mattered.
This is the single habit that separates people who get consistently good output from people who get occasionally lucky.
