Wan AI Prompts
A strong Wan prompt describes what the viewer can observe over time. Use this order: subject + action + camera + setting + light + sound + constraints. Remove any part that does not change the result.
Text-to-video formula
[subject] [observable action]. [camera framing and movement]. [environment]. [lighting]. [sound]. Preserve [important constraint]. Avoid [failure].
Example: A ceramic robot folds one sheet of blue paper into a crane. Medium close-up, one slow clockwise orbit. Clean wooden desk, warm morning window light. Soft paper folds and quiet mechanical ticking. Physically plausible hands, no captions, no logo.
Image-to-video formula
[what moves] + [how the camera moves] + [what changes in light/environment] + preserve [identity/geometry/background].
Example: The camera slowly pushes in while a narrow highlight moves left to right across the bottle. A few soft shadows drift behind it. Preserve the bottle shape, label geometry, and background composition.
Video edit formula
Transform [whole scene or named region] into [target treatment]. Preserve [timing, subject, camera, composition]. [audio instruction].
Example: Transform the entire clip into soft watercolor on textured paper. Preserve the original subject, timing, camera path, and composition. Keep original audio.
Three use-case recipes
Vertical social reveal
A compact product rises slowly from a dark pedestal as a thin rim light turns on behind it. Vertical 9:16 composition, slow push-in, high contrast, five-second reveal, no text, no logo distortion.
Cinematic B-roll
Close-up of rain collecting on a train window while distant city lights stretch into soft bokeh. Static camera with subtle handheld breathing, cool night tones, natural rain ambience, no people, no captions.
Artwork animation
Clouds drift gently from left to right, water ripples outward in two slow rings, and foreground grass moves in a light breeze. Preserve the illustration style, palette, character silhouette, and frame composition.
Debug one variable at a time
If the camera is wrong, change only the camera sentence. If anatomy fails, reduce action complexity and use a source image. If the subject drifts, repeat the preservation constraint once; repeating it throughout the prompt usually adds noise.
Frequently asked questions
Should a Wan prompt include a negative prompt?
Use explicit constraints such as no captions, no watermark, or preserve label geometry when the selected endpoint supports them. A short targeted negative instruction is easier to debug than a long generic list.
Are longer prompts better?
Not automatically. More words can create competing actions and styles. A complete shot plan in one paragraph is generally more controllable than a long mood board.
Can I mention a real person?
Only use people and media you have the rights and consent to use. Deceptive impersonation and explicit high-risk requests are blocked before provider spend.