How to write prompts for AI video
An AI video model only knows what you tell it. The clearest prompts read like a short description of a single shot in a film: who is in it, what they do, where, and how the camera sees it.
A structure that works
- Subject: who or what the shot is about. "A lighthouse keeper in a yellow raincoat."
- Action: one clear thing that happens. "Climbs the spiral stairs."
- Setting: where and when. "At dusk, as a storm rolls in."
- Camera: how it moves. "Slow push in", "tracking shot", "aerial", "handheld".
- Light and mood: "warm lamplight, rain streaking the windows".
- Sound, on models that make it: music, ambience, and spoken lines in double quotes.
Before and after
Vague: "a lighthouse in a storm".
Clear: "A lighthouse keeper in a yellow raincoat climbs the spiral stairs at dusk as a storm rolls in. Slow push in, warm lamplight, rain streaking the windows."
What to avoid
- Too much at once. Five seconds fits one action, not five.
- Saying what you do not want. "No people" can put people in. Describe what you do want instead.
- Real people, and other studios' characters. Studio does not make them, and the video services refuse them.
Experiment cheaply
Try a prompt on a lower-cost model at 480p first, such as Seedance 2.0 Mini or Wan 3.0. New accounts get 50 free credits, enough for one of those. When the prompt works, make the final version on the model and quality you want. In Studio, Reuse settings on any clip puts its prompt back in the form so you can change one thing at a time.