Cinematic Scene Case Library

Use this skill as an optional, low-permission retrieval layer for cinematic scene structure. It supplies reusable scene patterns and filtered handoff facts; it does not generate images, videos, media bindings, or final Seedance/H3 Prompts.

Retrieve only when useful

Retrieve a case when the request is abstract, lacks scene structure, asks to repair a Prompt, or explicitly asks for a case reference. Skip retrieval when the request is already concrete, shootable, and complete enough for the owning director skill. Load the smallest relevant set, normally one case from the index.

Workflow

1. Classify the request into one family in the index: action, performance, dialogue, environment, camera, physics/continuity, or mixed scene.
2. Read the selected case file and extract only the model-neutral pattern, variable slots, portable constraints, and the handoff fields needed by the next owner.
3. Treat the Prompt score and all provenance as Prompt-only audit evidence. It is not evidence of rendered-video quality, and no source media was inspected.
4. Apply user-locked facts and target-model rules first. If a case conflicts with either, report the conflict and omit the conflicting suggestion.
5. Return a filtered guidance package using guidance-package-schema.md, not a copied case file and not a final Prompt.

Ownership and isolation