How to write AI video prompts that actually render well
Most disappointing AI clips come from vague prompts, not weak models. A prompt that names one subject, one action, one camera move and one lighting idea renders far cleaner than a paragraph of competing ideas.
The five-part formula
Write your prompt in this order and you will rarely get mush: subject, action, camera movement, lighting, look. Each part is short, and none of them contradicts another.
- Subject — one clear thing: 'a lone diver', 'a matte black bottle'
- Action — what it does: 'drifting', 'rotating slowly'
- Camera — how the shot moves: 'slow push-in', 'low orbit', 'static wide'
- Lighting — 'volumetric light', 'golden hour side light', 'cold studio key'
- Look — 'cinematic 35mm grain', 'clean product render', 'anime cel shading'
A worked example
"A lone diver drifting through a bioluminescent reef at midnight, slow push-in, volumetric light, 35mm grain." One subject, one action, one move, one light, one look — under twenty words, and it renders consistently.
Mistakes that ruin a render
- Two subjects doing different things in a five second clip
- Cutting between shots — a single generation is a single shot
- Asking for readable text or logos; AI video rarely spells reliably
- Fast, complex camera moves in short durations
- Piling on ten style words that pull in different directions
Iterate one variable at a time
If a clip is close but not right, change only the camera move, or only the lighting. Changing everything at once makes it impossible to know what helped. Keep the prompts that worked — they become your house style.
When to start from a photo instead
If the subject already exists — your product, your artwork, your storefront — skip prompt writing. Upload the image, let expidvid scan it, and edit the motion description it produces. The photo anchors the first frame, so the result looks like your asset rather than an approximation of it.