Anatomy of a video prompt: the parts the model reads, in the order it reads them

People arrive at video generation from image generation, and they bring the habits with them: a long stack of style words, a list of quality claims, a mood. Then they add 'camera slowly pushes in' at the end and wonder why the result drifts, morphs, or simply sits there vibrating.

A video model is being asked a fundamentally different question. An image model has to answer what does this look like. A video model has to answer that and what changes between the first frame and the last — and the second question is the one that decides whether you get a shot or a mess.

Almost every reliable improvement in video prompting comes from separating those two questions and answering them in a fixed order, rather than mixing them into a single paragraph and hoping the model untangles it.

Write every prompt in five ordered blocks. The order matters because these models weight the opening of a prompt more heavily, and because later blocks are read as modifying earlier ones.

1. Subject — what is in the frame, described physically. Materials, colours, scale, position.
2. Action — what the subject does across the duration. One action, not three.
3. Camera — what the lens does. Static, or one move with a direction and a speed.
4. Light — direction, quality, colour. This is also what fixes the look between frames.
5. Constraint — what must not happen. The shortest block and often the most load-bearing.

A prompt written this way is longer than most people's, but almost none of it is style vocabulary. It reads like a shot description on a call sheet, which is exactly what it should be.

Write 'she picks up the cup, drinks, and sets it down' and you will get an incoherent five seconds in which none of those three things fully happens. The model does not have a shot-planning stage; it interpolates between a beginning and an end. Give it three beginnings and three ends and it averages them.

The rule is one action per generation. If the shot genuinely needs three, that is three generations and an edit — which is also how it would be shot with a camera.

The same logic applies to the camera. One camera move per shot. A push that becomes a pan is two moves, and the transition between them is where models produce their worst artefacts.

Terms that reliably work, because they describe a physical movement with a direction:

• Slow push in / pull out — the camera translates along its axis. Reliable across every model.
• Pan left / right, tilt up / down — rotation from a fixed position. Reliable.
• Static locked-off shot — the strongest instruction of all, and underused. If you want no camera movement, say so explicitly; silence is not read as 'static'.
• Handheld — adds motion, not shake specifically. Use when you want life, not when you want instability.

Terms that mostly fail:

• Dolly zoom, vertigo effect — requires two coordinated changes. Almost never renders correctly.
• Crane, jib, drone shot — often interpreted as 'aerial view' rather than as a move, changing your framing entirely.
• Cinematic camera movement — a mood word wearing a technical costume. It adds nothing and takes up space.

Add a speed to any move you request. 'Slow' and 'very slow' are the two you will use most; unqualified moves tend to come out faster than intended and faster movement is where morphing appears.

The last block names what must not happen. Not a generic quality list — the specific thing this particular shot is likely to get wrong.

If the subject is a product, the constraint is that the product must not change shape, colour or proportion. If the subject is a person, it is that facial features must stay consistent. If the frame has text in it, it is that the text must not change.

The bottle must not change shape, label position or colour.
No additional objects enter the frame. No cuts.

'No cuts' is worth including almost always. Video models will happily invent a cut mid-generation, and the result is unusable in an edit where you needed a continuous take.

What does not help: long negative lists inherited from image prompting. Twenty exclusions dilute the three that mattered, and several models weight the negative field weakly enough that the list is mostly ignored anyway.

When a generation fails, the temptation is to rewrite the whole prompt. Resist it — you learn nothing, and the next failure starts from zero again.

Change one block, regenerate, compare. It is slower for the first hour and much faster forever after, because you accumulate real knowledge about which block controls which failure.

Keep a file of prompts that worked, organised by what they were solving rather than by project. Six months later the value is not the prompt itself but the record of which phrasing moved which behaviour — and that transfers to the next model, which the specific wording usually does not.