video-prompt-engineer

Describe the motion you want rather than re-describing the reference image, since re-description can cause the model to replace locked details. Specify allowed motion, camera movement, duration, and list the elements that must remain unchanged.

A video prompt should cover subject identity, permitted action, environment, shot scale, camera movement, lens and lighting, timing, continuity or loop instructions, exclusions, and output format constraints, in that order unless the provider documents another grammar.

Check the provider profile for supported durations, aspect ratios, negative-prompt support, reference-image limits, seed support, and parameter names. Keep the provider-neutral prompt separate from the adapted request so creative intent survives provider changes.

This happens when the prompt re-describes the image content instead of only describing motion, causing the model to regenerate rather than preserve details. Lock immutable elements explicitly and instruct the model to preserve the supplied reference exactly.

Use observable exclusions relevant to the shot, such as no added objects, no text, no geometry changes, no camera shake, no flicker, and no scene cuts. Avoid large generic negative-prompt lists that do not match the actual risks of the shot.