# How to write prompts that produce usable AI video

A video prompt has four jobs — subject, action, camera and style. Write them in that order, keep one idea per clip and iterate on the weakest element instead of rewriting everything.

Updated 2026-11-13. Also on the [web page](https://aivideogen.fyi/guides/how-to-write-video-prompts).

A good video prompt is a shot description, not a wish. The models that make usable clips respond to four ingredients, in roughly this order of importance: the subject, the action, the camera and the style. Write them deliberately and you will need fewer generations per finished second.

## 1. Subject first

Name the thing that must be in frame, specifically. "A stainless steel water bottle on a kitchen counter" beats "a product." If the subject is a person, describe only what matters — age range, clothing, action — and leave the rest to the model. Faces are where generations fail most often; keep people small in frame or out of frame when you can.

For product work, the honest path is image-to-video: generate or photograph the real product as a still, then animate it. A text prompt cannot reliably invent your actual product, label and all.

## 2. Action second

Say what moves and how. "The camera slowly pushes in" is an action. "It looks cinematic" is not. Useful action words: rotates, slides, pours, unfolds, rises, tracks alongside, settles into place. One clear motion per clip beats three competing motions. If the shot needs a hand pouring liquid and a label coming into focus, split it into two clips and cut them together.

## 3. Camera third

Camera language carries more weight than most beginners expect. Shot size (wide, medium, close-up), angle (eye level, overhead, low) and movement (static, pan, push-in, orbit) each change the output. Luma's Dream Machine page highlights camera-motion controls; Runway and Kling expose similar motion options on their generation pages. Learn each tool's vocabulary — the words that move the camera are the tool's own terms, not universal ones.

## 4. Style last

Style is the finishing layer: lighting, palette, film look, mood. "Soft morning light, neutral tones, shallow depth of field" tells the model more than "professional" or "viral." Keep style consistent across a series by reusing the same style sentence in every prompt. That consistency is what makes five generated clips cut together like one shoot.

## A template that works

> [Shot size] of [subject, specific] [action, one motion]. [Camera movement]. [Lighting and palette]. [Style references: film look, aspect ratio hints].

Example: "Medium shot of a matte black coffee mug on a wooden table, steam rising slowly. Camera pushes in gently. Warm morning window light, neutral wood tones, shallow depth of field. Clean product-ad look."

## Iterate on the weakest element

When a generation misses, do not rewrite the whole prompt. Find the element that failed and change only that: the subject got the wrong shape, sharpen the subject description; the motion was wrong, rewrite the action; the mood was off, adjust the style line. Each retry costs credits — Pika's pricing page puts a 5-second 720p clip at about 122 credits on Starter — so targeted fixes beat wholesale rewrites.

## Practical habits

- Generate three takes of the hero shot and pick the best; budget for it.
- Keep every prompt that produced a keeper in a document; they are your reusable shot library.
- Match prompt length to clip length: a 5-second clip needs one clear beat, not a paragraph of action.
- For anything with text in frame — prices, labels, headlines — plan to add the text in your editor afterward. Generative video garbles lettering reliably.

## Sources

- [Runway AI video generator (fetched 2026-11-13)](https://runwayml.com/ai-video-generator)
- [Luma Dream Machine (fetched 2026-11-13)](https://lumalabs.ai/dream-machine)
- [Pika pricing (fetched 2026-11-13)](https://pika.art/pricing)

## Questions

### How long should a video prompt be?

Long enough to pin down subject, action, camera and style — usually two to four sentences. Longer prompts dilute the important parts.

### Why do my generations look different each time?

Most models add variation on purpose. Fix the elements that matter with more specific wording, and expect to generate several takes.

### Should I write the prompt before or after I know the tool?

Write the shot list first. Each tool's prompt field takes the same four-part structure; only the syntax for camera moves and style tags differs.