How to reverse-engineer
a video prompt.
A prompt is a shot list with the camera left in. Start with the cut, the light and the movement — then add the subject. Here is the workflow, with a real example.
Start with the cut
The first useful fact about a reference is not “cinematic” or “moody”. It is where the shot starts and ends. A 0.8-second cut and a 4-second cut are different instructions even when the frame looks the same: one is a punctuation mark, the other is a hold.
Open the reference in the workbench and read the shot list. You want the duration, the aspect ratio and the number of shots before you write a single adjective. The cut list tells you whether the sequence is built on fast cuts, long takes, or one continuous move — and the prompt for one shot should not try to reproduce the whole sequence.
Measure before writing
Five things are measurable from the frame and the timeline, and all five belong in the prompt:
| Field | What it changes in the prompt |
|---|---|
| Duration and aspect | Whether the model is filling a 1-second beat or a 4-second hold, in 16:9, 9:16 or another frame. |
| Luma key and contrast | Low-key with deep shadows reads differently from high-key with soft fill. Say which one. |
| Color temperature and palette | Warm highlights and cool shadows, or a neutral product look. Three or four colors anchor the grade. |
| Motion direction | Locked off, slow push-in, pan right, handheld drift. A camera instruction is a decision, not a decoration. |
| Rhythm | If the cut lands on a 90 BPM beat, the shot may be built for that length. Say the duration; let the edit carry the rhythm. |
The workbench measures these and writes them into shots.csv and
prompts.md. You do not need to eyeball a palette or guess the light key.
The prompt skeleton
Write in the order a camera department would ask for the information:
[camera move], [shot size] of [subject] in [setting].
[light key], [contrast] contrast, [color temperature] palette.
Color palette: [#hex, #hex, #hex].
[style / texture]. [duration] s, [aspect].
No text, no logos, no watermarks, no extra limbs. The first line is the blocking. The second is the light. The third is the grade. The fourth is the finish and the technical container. The last line is the negative prompt.
This order also makes the prompt easy to edit. If the generation comes back too dark, change the light line, not the whole paragraph. If the movement is wrong, change the first clause.
A worked example
The demo reel’s second shot is a chrome ribbon on a cream plinth. The measured values are a high-key studio frame, cool highlights against a warm paper backdrop, a slow pull-out, and a 2-second hold in 16:9. A literal prompt reads:
Slow pull-out, medium shot of a liquid chrome ribbon sculpture on a cream stone plinth. High-key lighting, medium contrast, cool palette. Color palette: #d8c6a4, #8c1c0c, #1a1512, #f2ead9. Editorial product film, fine grain. 2.0 s, 16:9. No text, no logos, no watermarks, no extra limbs.
The sentence is not trying to be poetry. It fixes the camera, the light, the grade, the finish and the duration. The subject is the only part that came from the brief. That is the division of labour: measured grammar, human idea.
Three variants, three jobs
- Literal — reproduce the reference. Use it for a look test or a shot-for-shot study.
- Same camera, new subject — keep the grammar, swap the product or person. This is the most useful variant for an ad remake.
- Stylized — keep the camera and palette, reinterpret the medium. Useful when the reference is a mood, not a template.
The workbench generates all three per shot. Do not merge them into one mega-prompt; a prompt that asks for a literal remake and a stylized reinterpretation at the same time gives the model contradictory instructions.
Common mistakes
- Describing the subject and forgetting the camera. The model will choose a default camera, and it will usually be the most generic one.
- Stacking quality words. “Cinematic, 8K, masterpiece, award-winning” is not a shot description. Camera, light and duration are.
- Asking one generation to cover multiple shots. Split the sequence at the cuts; generate the shots that a single model call can actually hold.
- Trusting a motion label as a verdict. Pan and push-in are estimated from the pixels. If the estimate is wrong, override the camera in the brief.
- Forgetting the negative prompt. Text artifacts, logos and watermarks are common failure modes; say no to them explicitly.
The prompt is a scaffold, not a guarantee. A good generation still depends on the model, the seed, the reference image and your taste. What the scaffold buys you is a consistent starting point you can iterate from.
Questions
Do I need a vision model to write a video prompt?
No. A useful first draft needs the cut length, aspect ratio, light key, palette, motion and rhythm, plus a sentence from you about the subject and setting. A vision model helps when you want a semantic description of the subject; it is not required for the measurable half.
Why does my prompt still look generic?
Most generic prompts are missing a decision, not an adjective. Name the shot size, the camera move, the direction of the key light, the color temperature and the duration. If those are vague, the model fills the gaps with its own defaults.
Should I paste the palette hex codes into a video prompt?
They are useful as a color anchor, especially when you want a specific grade. Keep them to three or four colors and describe their role — warm highlights, cool shadows, a neutral background — rather than listing ten codes.
Can I use the extracted frame as an image-to-video reference?
Yes. Every shot in the pack includes a first frame and a mid frame. Use the first frame when your model supports image-to-video; use the mid frame when you want a representative still for a look test.
How long should a shot prompt be?
Long enough to fix the decisions and short enough to stay coherent: usually 60–120 words. A prompt that lists twenty adjectives without a camera, a light direction or a duration gives the model more room to wander, not less.
Try it on a clip.
The workbench writes the measured half; you supply the idea.