video-gen

Video generation

This skill orchestrates four published flows:

• C0 local concat — video_compose joins compatible existing MP4 clips.
• Timeline render — video_compose locally turns images/screenshots and existing video clips into a video with overlays, motion, transitions, soft or burned subtitles, source audio, and optional BGM. Narration is an optional network feature.
• Single AI clip — one paid video_generate call.
• Shot-book AI film — author a shot book, generate frames with image_generate, then ONE paid video_render call renders and stitches.

The AI video model is fixed by pi-video-gen.defaultModel; generated images use pi-image-gen's active model.

A. Workflow rules

1. Route first. Choose the right flow before anything:

   User goal	Flow
   Existing local mp4 clips to join	video_compose (C0 — lossless, local, no paid models)
   One AI-generated moving shot	video_generate
   Multi-shot film with keyframes	video_render
   Promo/explainer from images, screenshots & clips	video_compose (TimelineSpec — local mixed-media render; optional network TTS)

2. Local flows stop here. For C0 follow §A0; for Timeline follow §A1. Do not run the AI preflight, shot-book steps, or paid confirmation gates below. Timeline only needs image_generate when its source images do not already exist.

3. AI preflight only. For video_generate or video_render, call video_capabilities and respect the active model's duration range and audio support, frame support, and trusted asset modalities. Confirm image_generate is available only when source frames need to be generated (/video-gen doctor checks; config health is /image-gen list).

4. Pick the AI flow. A vague idea or a script that needs multiple shots → shot-book flow below. One moving shot → follow §A4 and skip the shot-book stages. A still → image-gen.

5. Write the shot book in conversation (shot-book only) (schema in §B). If the user only has a vague idea, first be the screenwriter: three-act structure, filmable actions ("show, don't tell"), concrete visual detail. Iterate with the user in chat.

6. Confirmation gate 1 (shot-book only). Show the shot-book summary — shot count, character list, estimated image calls (~2N+3C) and video calls (N) — and get an explicit go-ahead. Default small: 1 scene, 3–5 shots unless the user asks for more.

7. Source stage (shot-book only). Use trusted portrait assets per §A2 when Seedance will receive a recognizable real person. Otherwise generate the required character portraits and per-shot frames via image_generate per §C. Show each generated batch to the user.

8. Paid confirmation. Before video_generate, confirm its one paid call. Existing explicit approval covers the unchanged request; reuse it instead of asking again. For a shot book, once frames are ready, state "about to make N paid video calls" and get an explicit render order. Then assemble the render spec and call video_render ONCE.

9. Cost honesty. AI video calls are paid and take minutes each. Never state amounts (prices change); state call counts and durations.

10. Revisions. The render spec is immutable per job directory. Text-stage revisions happen in chat (regenerate frames as needed); a revised film goes in a NEW job directory. NEVER suggest "delete shots// and rerender" — that breaks downstream dependencies. Rerunning the SAME spec path resumes an interrupted job (finished shots don't re-bill).

11. Degradation negotiation. If video_render preflight fails (e.g. last frame unsupported), present the options (switch model / edit spec / allowDegradations) and let the user choose. Never degrade silently. When the model's nativeAudio is false, omit audio cues and obtain acceptance of silence if audio was requested.

12. Cancellation honesty. Interrupting stops local polling only — remote tasks may keep running and billable (Ark cancellation is unverified). Say so.