| name | dircreative-longform-reference-planner |
| description | Build global and per-sequence reference pack contracts for longform AI video workflows. |
Longform Reference Planner
Required Knowledge
docs/film-preproduction/longform-decomposition-policy.md
docs/film-preproduction/capability-aware-generation-policy.md
docs/film-preproduction/reference-locking-policy.md
docs/film-preproduction/schemas/longform-reference-pack.yaml
docs/film-preproduction/schemas/reference-pack-manifest.yaml
docs/film-preproduction/research/model-reference-behavior.md
Inputs
- sequence plan
- visual bible
- reference pack manifest
- image prompt manifest
- target model list
Outputs
- longform reference pack
- global reference pack
- per-sequence reference pack plan
- clean start/end frame requirements
- external generation instructions
Rules
- Separate global identity/style assets from sequence assets.
- Use sequence packs as the normal unit for 180s work.
- Do not make one clean image per shot unless QA or model constraints require it.
- Every asset must have one primary job and one
asset_output.status.
prompt_only must remain fully useful without generated files.
assisted_generation requires recorded user authorization.
external_generation must include upload/import instructions.
- Resolve every target through
capability_card_id + version + provider_surface; never emit family aliases or family-keyed policy maps.
- Copy no numeric reference limit into the pack; validate limits from the exact current card at audit time.
- Keep dense boards out of direct frame slots whenever the exact card requires a clean image input.
- Use clean first/end frames only for exact cards whose reference modes authorize those inputs.
- Do not generate images or videos.
skill_run_receipt
Record global packs, sequence packs, output statuses, user gates, exact-card direct input policy, QA status, and next_recommended_skill: image-prompt-compiler.