DIRcreative

Plans and prompts are not media.

Invocation Boundary

Start only on explicit DIRcreative invocation (for example, $dircreative) or validated adco.specialist-exchange selecting dircreative.film-preproduction. DIR/ADCO source maintenance stops runtime execution (source_maintenance); runtime tests use isolated fixtures. Never modify project instructions to activate this Skill.

Router Contract

Choose the route by unfinished work. Read exactly one selected Route Card and its active reference together; replace references by stage and reuse loaded bodies. There are zero unconditional protocol reads. An obvious Fast or Studio request needs no router tool call. Canonical asset work still uses the stage selector before its first image.

Full film preparation starts in Studio even when images are authorized. Delivery executes already-designed assets after the required film craft.

Requested result	Mode / route	Task reference
Bounded copy/script edit	Fast / copy_revision	references/copy-script.md
One shot / small storyboard review	Fast / shot_optimization or storyboard_review	references/shot-storyboard.md
Bounded prompt edit	Fast / prompt_revision	references/prompt-model.md
Two-or-more-person blocking, movement or reverse-shot discussion	Studio / film_development with spatial_discussion layer	references/spatial-discussion.md
Other bounded revision	Fast / bounded_revision	none
Reference-video analysis	Studio / video_distillation	references/video-distillation.md
Client-readable story only	Studio / film_development	references/client-story.md
Full film/scene preparation, including real images	Studio / film_development	references/film-development.md
Recurring character master/derivative	Studio / film_development	references/character-master-sheet.md
Execute an already-designed asset / client delivery	Delivery / generation_authorization or client_delivery	references/generation-delivery.md
Validated ADCO handoff	Delivery / adco_specialist_exchange	references/specialist-exchange.md plus descriptor schema

Route Cards: routes/fast-task.md, routes/studio-development.md, routes/delivery-audit.md. Run scripts/dircreative_route.py only to validate an ADCO handoff or resolve genuine ambiguity; it is not a mandatory creative preflight. A handoff must validate its actual descriptor, sources, locks and output scope.

Work from the requested outcome

Preserve supplied material and the whole requested output set. Story-only stops at story; full preparation continues through script, shots, assets and prompts.

• Fast: one bounded change, zero Threads or Director Room. Return the edit.
• Studio: one controller, at most three dynamic professional perspectives and one critical pass. Establish dramatic cause, action, visual/sound choices and continuity before matrices.
• Delivery: check tool/reviewer/readback availability early; validate the asset and dependencies, execute authorized work, inspect, then continue.

Named 导演组/创意组 use the active craft card. For real/ambiguous collaboration read docs/film-preproduction/director-room-routing.md. Distinguish perspectives, subagents and user tasks. Never simulate delegation. No nested dispatch through ADCO; disclose unavailable tools before substituting single-agent work.

Project Interaction

For a new story or film-preproduction project with unknown preference, use the host's native choice control for 讨论共创（推荐）、直接执行、关键节点讨论 and wait for a real reply. 自行编排、自主完成、直接执行 are direct-mode choices: record them for this project and do not ask again. 创意组 covers story, motivation and writing; 导演组 covers performance, shots, space and sound. They are perspectives, not a subagent claim. Detailed tool, deduplication and fallback rules live in docs/film-preproduction/chat-co-creation-interface.md.

Skip this setup for bounded edits, analysis, already-designed assets, source maintenance and ADCO workers. Discussion answers the current issue without a mandatory question each turn. When a direction is ready, state direction, style and concrete scope, then offer 按此开始制作 or 继续调整. Production start only executes that shown, authorized scope; it never authorizes unlisted media or client delivery.

Intelligent Skill Stack

Infer required assets from the requested deliverables and current story, including “九宫格分镜和相关资产”. Apply references/shot-development.md through select --stage <stage> --format task. Recurring human assets default to one identity per master: frontal portrait plus front, left, right and back full bodies; read references/character-master-sheet.md before designing them. A standalone portrait or text-only board stays bounded. Initial assets need design, not their own future images.

For canonical asset generation, use references/image-execution.md: prepare the role-bound Jingzao spec, validate the current packet, submit its returned arguments, record the real PNG, and inspect the batch before an actual downstream dependency uses it. Repair failed candidates within scope. A generic compiled prompt, later plan, or unchecked generation cannot replace this sequence. Contact/action rehearsals and narrative frames retain their selected craft and coverage requirements.

Reuse the scenario/gaps/staged-pass selector: one craft owner, needed collaborators and at most one validator. Read each selected method once and reuse it; when a public helper already returns the needed output, do not read its source or tests. “已用” requires applying the selected body.

External User Gates

External gates are concept_lock, generation_authorization and client_delivery_approval. Reuse scoped authorization; image, video and delivery remain separate. Unknown facts block only affected claims.

State and truth

standalone_chat owns response/state. In orchestrated_worker, ADCO owns adoption, versions, visibility and cleanup; return artifacts, domain QA, status and open questions. Keep one creative source and one physical owner for identical bytes; mark only affected descendants stale. Use runtime/state-snapshot.schema.json for resume or multi-file execution; follow references/project-hygiene.md after the first useful artifact.

Quality and completion

Preserve source truth, project vocabulary, voice and genre. Before affected assets or prompts, reconcile the current source and adjacent continuity; proposed revisions do not become source facts. Keep offscreen presence distinct from an actual exit. Use references/copy-script.md for language craft; load references/humanization-workflow.md only for requested layered diagnosis, demonstrated document-scale defects or explicit providers.

Keep Prompt IR model-neutral; compile one verified model surface at a time. Preserve reference roles, acting, scene/support truth and preserve/change limits. Record final candidates in ai-film-production-ledger; it neither executes nor approves.

Whole-film coverage follows references/film-development.md. A ready root asset starts as soon as its design/authorization are ready; do not wait for unrelated downstream coverage. Generate independent assets in a batch; review only blocks the real next dependency. Preserve five-view masters, approved parent locks and actual image QA. Required coverage remains checked at its own stage and delivery gate. Show a newly completed story/shot artifact briefly with its openable file; preview each saved image immediately, but do not call either acceptance. Continue authorized self-check and repair without stage-by-stage process broadcasts. visual_assets_complete requires real canonical PNGs, visual review and trusted readback across the requested scope. It is not a finished video or user/client acceptance. Required pre-video gate gaps remain work to finish, not completion.