How to write an AI video prompt that directs a shot
Treat the prompt as a compact shot brief. Prioritize what must move, how the camera observes it, and what should stay stable.
Updated August 17, 2026
1. Start with the shot's single job
Choose one primary action and one visual outcome. A prompt that asks for an entrance, transformation, product demo, crowd reaction, and aerial reveal in the same short clip gives the model competing priorities.
- Name the subject
- Use one clear action verb
- State the intended framing or publishing format
2. Add information in a stable order
A reliable order is: subject and action, environment, composition, camera movement, lighting, pace, and sound. This makes the prompt easier to revise because each phrase has a job.
- Subject: who or what viewers follow
- Action: what visibly changes
- Camera: static, push-in, orbit, pan, track, handheld
- Light: source, direction, softness, color
- Sound: ambience, dialogue intent, or silence where supported
3. Use camera language that can be seen
Concrete direction such as “slow lateral tracking shot at waist height” is more actionable than “make it cinematic.” Combine one camera movement with framing and pace; avoid stacking four movements into a five-second clip.
- Static locked shot for product accuracy
- Slow push-in for emphasis
- Side tracking for walking action
- Orbit for dimensional product reveals
4. Protect what should not change
For characters and products, explicitly preserve the critical identity, logo, wardrobe, shape, or typography. Keep the constraint list short and specific.
- Preserve facial proportions
- Keep package text unchanged
- No new objects or extra limbs
- Maintain the source-image composition
5. Iterate one variable at a time
Save a useful result, then change only the motion, camera, duration, or model. If every setting changes at once, you cannot tell which choice improved the shot.
- First pass: subject and action
- Second pass: camera and pace
- Third pass: light and sound
- Final pass: negative constraints only where needed
Ready to put the workflow into practice?
Open the Cdance generator