Updated · 7 min read
How to prompt AI video: camera, light, motion, sound
Video prompts fail in predictable ways: too many ideas, no camera, no time order, and words the engine paints on screen. This guide is the structure we use for every render, with the fixes for the failures we see most.
The structure
- Realism anchor first if you want realism: photorealistic live-action, not animation, not cartoon.
- Subject and setting in one sentence.
- Camera: one move, named. Slow push in. Handheld follow. Static wide.
- Beats in time order: what happens first, then next, then last. Say the action is already happening if you want motion from the first frame.
- Sound: dialogue in quotes, effects in angle brackets, music in parentheses when the studio supports the notation, or just described in words.
- Light and lens: golden hour, overcast, 35mm, shallow depth of field, film grain.
- Consistency line: keep the face consistent with Image 2, keep the product consistent with Image 1.
What to leave out
- Hashtags, brand slogans and anything you do not want drawn as on-screen text.
- Named real people. The platform refuses celebrity likenesses.
- Three ideas. One shot, one idea. Make a second clip for the second idea.
- Timestamps and shot lists on single-take engines. Describe one continuous move instead.
Dialogue that lands
Keep spoken lines short: about fourteen characters per second of screen time. Put each line in quotes and attach it to the person speaking. On Kling the words render as speech; long sentences get rushed. If the clip is a testimonial, one honest line beats a paragraph.
Realism
Seedance 2.0 drifts to animation when a prompt says anime, action or stylised and has no references. Anchor it with a photo reference and the realism line at the top. Add texture words: pores, real fabric, no plastic skin. Veo 3.1 and Kling 3.0 hold realism with less coaxing.
Common failures and fixes
- Opens on a still frame: write that the action is already in motion from the first frame.
- Text on screen is gibberish: shorten it, move it to a physical surface like a screen or sign, or use Seedance rather than Kling for long strings.
- Face changes between shots: cleaner frontal reference, fewer competing references, or a saved character.
- Product label wrong: flatter product photo as Image 1, and a consistency line in the prompt.
- Music sounds like a real song and the render fails: the platform retries without the score. Describe original ambient music rather than naming a genre with a famous sound.
Let Director Mode write it
In the video studio, Director Mode expands a rough idea into a structured prompt in this shape for 5 credits. In Image Studio, Enhance does the same for stills for free and knows a still is one frozen instant, not a sequence.
Questions, answered
How long should a prompt be?+
Long enough to name the subject, the camera, the beats, the sound and the light. Usually 60 to 150 words. More ideas do not help; more precision does.
Does the engine see the whole prompt?+
Yes, plus safety guidance the platform adds and never stores on your render. Your library shows exactly what you wrote.
Can I reuse a prompt?+
Every render in your library has a copy-prompt button, and community posts can be remixed into your studio with the prompt loaded.