direct-video-creation

Lead with subject and identity, then describe one principal action, environment, camera move, light, and an explicit final state. Keep beats proportional to duration and use concrete physical direction rather than abstract mood words.

Use the image-to-video method: assign the still a clear authority role such as identity or composition, describe only the motion delta rather than redescribing the image, and state the intended final state. Use a clean approved reference to reduce drift.

First/last-frame mode interpolates motion between two authoritative endpoint images. The last frame must be generated as a controlled edit of the first frame, and the video prompt describes only the physical transition between them.

Only when the inspected model capability explicitly supports timestamped multi-shot generation. Otherwise create one independent generation node per shot and return an ordered sequence manifest with continuity handoffs between clips.

Chaotic motion usually comes from too many beats or camera moves in one prompt; keep one principal action and one primary camera move. Identity drift typically means a weak reference or excessive transformation distance, so use a cleaner still and reduce the change.

Audio support depends on the inspected model capability and must never be inferred from a model family name. When native audio is unsupported, produce a separate audio cue sheet aligned to the generation timeline and mark synchronization as external work.