# How do you write good prompts for AI video?

**Short answer:** Describe one shot as a camera would see it: who is in frame, one action, the setting, a concrete light source and the framing, with style rules set once for the whole project. State what you want rather than what you do not want, and keep colour words off skin.

URL: https://arc.katechon.technology/answers/how-to-write-prompts-for-ai-video
Updated: 2026-10-18

## A prompt structure that works

One shot, field by field:

| Field | Example |
| --- | --- |
| Framing | Medium close-up, eye level |
| Subject | Maren, the herbwoman (reference attached) |
| Action (one) | She wraps a bandage around the boy's arm |
| Setting | Her cottage, night, snow at the window (reference attached) |
| Light | A single oil lamp on the table, warm, low |
| Sound or line | She says quietly: "Hold still." |
| Length | 6 seconds |

## Rules from our productions

- **One action per shot.** Two actions usually means one of them is skipped or merged.
- **Name a concrete light source.** "A single lamp" beats "moody lighting".
- **Colours on the world, not on skin.** "Cold blue shade on her face" produced a blue-skinned character; put the blue on the walls.
- **State rules positively and where they apply.** "The red balloon is the only red" made the image model paint balloons into set sheets. "No red in sets or wardrobe" plus "empty plate, no balloons" in set descriptions fixed it.
- **Tag continuity in every shot.** Helmets flipped between cuts until each shot stated helmet state.
- **Name who speaks.** Unattributed dialogue can come out as a narrator's voice over the wrong shot.
- **Avoid famous-face archetypes.** A description that reads like a well-known actor can get shots blocked as a lookalike.
- **Put style in the project, not the shot.** Medium, palette and lens apply to every frame; restating them per shot invites drift.

## Prompts are not enough for a series

However good the prompt, the model reinterprets words every time. For anything with recurring characters, attach reference images; see [text, image or reference to video](https://arc.katechon.technology/answers/text-to-video-vs-image-to-video-vs-reference-to-video). Rewriting a shot's prompt cost us more than recasting a character's reference when a face kept getting blocked.

## Iterate cheaply

Test a prompt on a storyboard frame (13 credits in Arc) or on Seedance 2.0 Mini (47 credits for 5 seconds at 720p) before rendering the final take on a pricier model. Arc can preview the final prompt for any shot before generating it.

## Where Katechon Arc fits

Arc writes the per-shot prompt from the script plan, references and project style, and lets you preview and edit it before generation.

Arc's prompts target Seedance 2.0 models; prompt conventions for other models' own apps differ, and Arc does not offer them.

## Related answers

- [How does text-to-video AI work?](https://arc.katechon.technology/answers/how-does-text-to-video-ai-work): A text-to-video model is trained on huge numbers of video clips paired with descriptions, and learns to turn random noise, step by step,…
- [How do you make AI video look less like AI?](https://arc.katechon.technology/answers/how-to-make-ai-video-look-less-ai): Pick one visual language and hold it everywhere, keep faces and places consistent with references, and put colour and lighting words on the…
- [How do you keep costumes and props consistent in AI video?](https://arc.katechon.technology/answers/how-to-keep-costumes-and-props-consistent-in-ai-video): Write the state of every costume and prop into every shot (helmet on or off, sword in hand or sheathed), give each distinct look its own…
- [Why do AI video generations fail or get blocked?](https://arc.katechon.technology/answers/why-do-ai-video-generations-fail-or-get-blocked): Three different things look like a failure: transient job errors that pass on retry, content moderation that blocks an input or an output,…
- [How do you turn a script into a video with AI?](https://arc.katechon.technology/answers/how-to-turn-a-script-into-a-video-with-ai): Break the script into shots, give every character and location a reference image, storyboard each shot, generate one clip per shot and…