空中群集｜近表面转向到稳定落地
Create a new execution from the reusable mechanism, not a surface reskin of any reference video.
Workflow
- Read references/summary.md for fit, input slots, and the concise starting example.
- Read references/template.md completely before drafting.
- Preserve the user's subject and objective, then fill every missing scene, trigger, event-chain, camera, sound, and payoff slot.
- Treat these anchors as mandatory:
- 领衔主体始终可追踪
- 宽景重建方向和深度
- 最后接触面与余动可见
- Change source-specific people, setting, props, palette, dialogue, shot boundaries, sound family, and payoff surfaces together.
- Compile MiniMax H3 and Seedance 2.0 independently; do not leak one model's field syntax into the other.
- Silently verify anchor coverage, continuity, timing, physical causality, and anti-copy distance before returning the final prompts.
MiniMax H3
- Use the current official H3 task grammar and language profile.
- Separate integrated visual action, overall soundscape, and non-diegetic music or explicit silence.
- Use timestamps only to make event order executable; keep shots consecutive and duration-safe.
Seedance 2.0
- Define stable subjects once and preserve event order, camera motivation, sound intent, and overall duration.
- Use a simple paragraph for simple concepts or an ordered multi-shot structure for complex concepts.
- State overall duration once; do not copy H3 field labels or per-shot timecode grammar.
Output
Return one concept sentence, the filled slots, an anchor-coverage check, one MiniMax H3 prompt, one Seedance 2.0 prompt, and a short anti-copy validation note.
Non-Negotiable
- A subject-only request is underspecified. Preserve that subject and complete the mechanism instead of returning a generic portrait or showcase.
- Do not omit, reorder, or weaken a required anchor.
- Do not reuse source people, brands, dialogue, exact storyboard, signature props, exact timings, or recognizable effect surfaces.