How to Write Prompt for AI Video Generator
Learn how to write a prompt for an AI video generator with a five-part formula, before-and-after examples, notes per generation mode, and a review loop.
To write a prompt for an AI video generator, describe five things in plain words: the subject and its action, the setting and light, the camera, the style and mood, and the sound or dialogue. Give the clip one clear action, say what must stay unchanged, and keep spoken lines short. That formula works across models, from Seedance 2.0 and Seedance 2.5 to image-to-video clips made from a still in VideoGen.
What a video prompt needs to say
If your prompt leaves open what moves, where it happens, how the camera sees it, or what it sounds like, the model guesses.
A good prompt answers four questions:
- Who or what is on screen, and what are they doing?
- Where are they, and what is the light like?
- How is the shot framed, and does the camera move?
- What does the viewer hear?
The AI video generator page describes the VideoGen workflows these prompts feed into.
The five-part prompt formula
Write the prompt in this order:
- Subject and action. One subject, one visible action with a physical cause.
- Setting and light. Place, time of day, light source, and weather if it matters.
- Camera. Shot size, angle, and one camera move or a locked shot.
- Style and mood. Color, texture, pace, and a reference style if it helps.
- Sound and dialogue. Ambience, key sound effects, and any spoken line in quotation marks with a delivery note.
One or two sentences per part is enough. Longer prompts tend to add contradictions, not detail.
Before and after: three rewritten prompts
Each pair shows a vague prompt and the same idea written with the formula.
Before: A woman drinking coffee, cinematic, beautiful lighting.
After: A woman in a gray sweater lifts a white mug and takes a slow sip at a kitchen counter. Morning window light from the left, soft shadows. Medium close-up, camera holds still. Warm, calm color, shallow focus. Quiet room tone, a light clink as the mug touches the counter.
Before: Product video of a running shoe, make it exciting.
After: A black running shoe lands heel first on wet asphalt and water sprays outward. Night street, blue neon reflections, light rain. Low angle, slow push in toward the shoe as it lands. High contrast, crisp detail. Rain patter and one sharp splash on impact.
Before: A chef talking about his restaurant.
After: A chef in a white jacket leans over the pass, points at a plated dish, and says, "Table seven, let's go," in a calm, confident voice. Busy kitchen, warm overhead light, steam in the background. Medium shot, slight handheld movement. Documentary look. Pans clatter and a ticket printer clicks under the line.
Prompts for different generation modes
The formula holds across modes. The emphasis shifts.
Text-to-video clip
Use all five parts. For Seedance 2.0 and Seedance 2.5, put spoken lines in quotation marks, describe how they are delivered, and keep dialogue short. Numbered shots give a multi-shot sequence in one generation. The Seedance prompts page has copy-paste examples for action, dialogue, product ads, and multi-shot stories, and the Seedance 2.5 prompting guide covers longer clips with more reference material.
Match the dialogue to the clip length. In VideoGen, a single clip can run from 4 to 30 seconds on Seedance 2.5 and from 4 to 15 seconds on Seedance 2.0 Mini, and VideoGen moves to a model that supports the requested length rather than shortening the request. In a storyboard scene, the spoken script may fill the whole clip.
Image-to-video
The image already fixes the subject, setting, light, and framing. Your motion prompt should describe only what moves and what stays the same:
Steam rises from the cup. The camera pushes in slowly. Keep the mug, the hands, and the background unchanged.
In a VideoGen project, "Convert images to videos" applies one motion prompt (or automatic motion) to every eligible still, with generated audio muted under a voiceover. The image to video page explains the options.
A whole project from a brief
Script-to-video and the prompt-to-video workflow take a script or a brief, not a shot prompt. Describe the goal, the audience, the key points in order, the tone, and the length. VideoGen builds the sections, and you refine single scenes with clip prompts afterward. For marketing-specific examples, read VideoGen prompts for marketers.
Avatar scenes
An avatar script is not a prompt. The script is what the avatar says, and the actor image decides the face and framing. Write the lines for the ear and keep each passage to one idea.
Common prompt mistakes
- Two or more actions in one clip. Pick one. Use a second clip or a numbered shot list for the next beat.
- Adjectives instead of events. "Dramatic" is a feeling. "The coat catches the wind as she turns" is an event the model can show.
- Dialogue that is too long for the clip. Describe the voice and the room instead of writing a paragraph.
- Forgetting sound. Models that generate audio need a description of it, or the sound may not match the scene.
- Nothing marked as fixed. If a product or face must stay the same, say so, and add a reference image when the prompt is not enough. In VideoGen, an image edit accepts up to 5 reference images, so you can fix the still first and then animate it.
Review and revise
Generate, then check the clip against four points:
- Motion: does the main action happen from start to finish?
- Consistency: does the subject stay the same shape, color, and identity?
- Audio sync: do sound and dialogue match what you see?
- End frame: does the clip end on a frame you can cut from?
Change one part of the prompt at a time. If the action is right but the shot is wrong, fix the camera line first. If the subject drifts, add a reference or state what must stay unchanged.
Frequently asked questions
How long should a prompt for an AI video generator be?
Long enough to cover the five parts in one or two sentences each, and no longer. Add words only when each one removes a decision from the model.
Can I put a spoken line in a Seedance 2.0 prompt, and how long can it be?
Yes. Put the line in quotation marks, describe how it is delivered, and keep it short enough for the clip, which runs from 4 to 15 seconds on Seedance 2.0 Mini in VideoGen. If the line needs more time, Seedance 2.5 clips can run up to 30 seconds.
Is a video prompt the same as a script for script-to-video?
No. A prompt describes one clip, while a script or brief describes a whole video that script-to-video turns into sections you can then refine clip by clip.
Do I need a different prompt for a Seedance 2.5 clip than for an image-to-video clip?
The five-part formula works for both, but the emphasis shifts. A Seedance 2.5 prompt covers all five parts, with dialogue in quotation marks and numbered shots for a sequence, while an image-to-video prompt describes only what moves and what must stay unchanged, because the still already fixes the subject, setting, and framing.
Start a project in VideoGen, write your first prompt with the formula above, and refine each scene on the timeline.
Merhannah A.
Richard D.