| name | dircreative-reference-image-planner |
| description | Plan reference image pack roles, board types, and layout specs before prompt compilation. |
Reference Image Planner
Required Knowledge
docs/film-preproduction/schemas/prompt-ir.schema.json
docs/film-preproduction/asset-intake-and-state-standard-v1.md
docs/film-preproduction/prompt-authoring-standard-v1.md
docs/film-preproduction/chat-co-creation-interface.md
docs/film-preproduction/chat-inline-visualization-interface.md
docs/film-preproduction/ad-reference-pack-generation-gate.md
docs/film-preproduction/research/storyboard-reference-analysis.md
docs/film-preproduction/research/image-prompt-style-system.md
docs/film-preproduction/research/model-reference-behavior.md
docs/film-preproduction/research/tapnow-agentic-canvas-lessons.md
docs/film-preproduction/reference-locking-policy.md
docs/film-preproduction/reference-consistency-gate.md
docs/film-preproduction/client-film-hard-gates.md
docs/film-preproduction/co-creation-gate-policy.md
docs/film-preproduction/longform-decomposition-policy.md
docs/film-preproduction/capability-aware-generation-policy.md
docs/film-preproduction/clean-frame-export-policy.md
docs/film-preproduction/sources/model-sources.yaml
docs/film-preproduction/schemas/visual-bible.yaml
docs/film-preproduction/schemas/reference-pack-manifest.yaml
docs/film-preproduction/schemas/longform-reference-pack.yaml
Inputs
- visual bible
- shot list
- asset requirements
Outputs
- reference pack plan
- reference pack manifest
- image layout spec
- board role map
- model-specific direct input policy
- exact capability-card resolution receipt
- reference rights gate
- asset output status map
- longform sequence pack requirements when target duration exceeds one generation unit
Chat Surface
Show the reference strategy as a user decision:
阶段: 参考图方案
智能体创作内容: number of boards/frames, each role, what each should contain.
逐镜头素材合同: for client-facing films, show whether each shot uses live action, UGC, reference video, existing Grok/ChatGPT/ImageGen image, prompt-only, or a new image prompt.
广告参考图组: product identity board, lighting/material/style board, storyboard/motion board, and clean frames for ad-film work.
不能直接喂给视频模型: list dense boards, labels, floor plans, storyboard boards, and direct-I2V risks.
镜头运动策略: show which board records camera movement, subject movement, timing, and transition logic.
用户确认点: ask whether to use the proposed pack, compress it, or split it further.
生成策略选择: show all-reference, hybrid, per-shot I2V, or minimal test when relevant.
模型能力依据: show exact capability card, version, status, provider surface, accessed date, and any official-source conflict; never show only a family alias.
权利状态: show which source, likeness, voice, brand/character, and music rights are verified, conditional, or blocked.
Keep generated media authorization as a separate later question.
Visual Decision Contract
Use skills/dircreative/assets/visualizations/stage-surface-registry.json#reference-asset-role-graph. Bind every asset node to ID, role, status, shot binding, and current hash-bound source; visually distinguish planning-only from direct-video-input roles and preserve the Mermaid fallback.
Rules
- Plan the fewest reference assets that still preserve readability and model safety.
- Resolve
capability_card_id + version + provider_surface from model-sources.yaml before naming duration, reference modes, audio, edit, extension, or upload slots. A family alias, latest, stale card, S4-only evidence, workflow-only card, or deprecated card cannot authorize generation.
- Preserve source tier and
accessed_on evidence in the manifest. If official sources conflict, keep the conflict visible and fail closed when the exact execution route remains ambiguous.
- Run the rights gate before proposing any direct image, video, audio, character, element, likeness, voice, brand, character, or music input. Unverified/blocked assets remain planning-only.
- Before any prompt or generation step for a client-facing film, create a per-shot asset/reference contract. Each shot must name source decision, source status, usage page or shot, planning/direct-input role, and blocker status.
- If the user says existing browser, Grok, ChatGPT, ImageGen, downloaded, or local images exist, complete browser/local intake before declaring missing assets or writing replacement prompts. If intake tooling is unavailable, record
TOOL_BLOCKED and the manual intake needed.
- Hero, acting, movement, or expression-critical shots must include a reference video, reference clip, or explicit motion note. Music/video sources must keep source, usage, and lock status so later versions cannot drop them.
- For assisted generation, lock visual truth in this order: character identity, scene geography/camera FOV, professional storyboard/motion page, then selected clean frames.
- Character consistency and scene consistency are mandatory. Do not batch-generate multiple boards that independently reinterpret the same character or location.
- Prefer a minimal V2 pack when drift risk is high: one character identity reference, one scene geography + camera FOV atlas, one professional storyboard/motion page, then optional clean first/end frames.
- Scene geography boards should be a single atlas when possible: panoramic or 360-degree environment, top-down route map, shot camera positions, FOV wedges, subject path, and vehicle path.
- Repeated character or scene content must inherit from the locked source asset; repetition must not introduce new face, wardrobe, prop, street, vehicle, or light-source facts.
- Shot count and image count are dynamic. Do not force three clean frames or one image per shot by default.
- Before planning frames, choose a reference strategy:
all_reference_sequence, hybrid, per_shot_i2v, or minimal_test.
- For fast 15-second work, allow more shots and more clean frames when the edit rhythm needs them.
- For all-reference models or modes, dense boards can be the main guidance if the prompt manifest records anti-misread clauses and accepted risk.
- For complex ad films with unresolved visual truth, propose a complete reference pack: identity, scene/FOV, optional lighting/material/style, optional storyboard/motion, and only the clean frames required by the selected model route.
- Do not treat a complete pack as a universal prerequisite. If the user supplies locked character/product and scene assets, compile the prompt from those assets plus written composition, camera, action, audio, transition, and conditional Look fields.
- A style/material/lighting board is optional when the look can be expressed precisely in the prompt and no repeated cross-shot drift is observed.
- A storyboard/motion board is optional when the shot list and Prompt IR already express timing, blocking, camera path, sound, and transition; it is a human-review artifact, not a mandatory model input.
- Do not generate a clean frame unless the selected model workflow requires a direct start/end image or a clean visual anchor.
- Default
visual_output_mode to prompt_only unless image generation is available and authorized.
- Record execution capabilities and
asset_output.status for every planned asset.
- For 60s, 90s, and 180s targets, create global references once and sequence packs one at a time in hybrid mode.
- Every reference asset needs exactly one primary production role.
- Every planned board must have a visible role label or manifest role label, such as character reference, scene + camera movement reference, professional storyboard + motion map, or clean first/end frame.
- A professional storyboard/motion page is mandatory once shot review is needed. Each shot cell must include timecode, duration, shot image region, detailed frame description, shot size, focal length, camera position, camera movement, subject blocking, sound, transition, and model risk.
- A professional storyboard/motion page defaults to
planning_only. A clean first/key/end frame is a separate, text-free asset; never imply a board crop is a clean frame.
- Use dense boards for human review, Seedance, and Veo only when the direct input policy allows it.
- Use clean first frames or start/end frame assets for Kling and Runway direct image-to-video inputs.
- For any direct first-frame/start-end-frame workflow, plan the required clean frames for that selected mode.
- Clean frames are separate assets or explicit clean-frame prompts, not hidden crops implied by a dense board.
- Do not mark direct I2V as ready when clean frames are only planned; keep
asset_output.status as prompt_ready, external_pending, or generated_candidate until the user locks real images.
- Separate planning boards from direct model inputs when labels, arrows, floor plans, or panel borders could be misread.
- Add
user_decision_gate before locking visual direction, reference pack plan, identity/product references, environment/camera path, storyboard motion, and clean first frames.
- Record the matching co-creation gate before treating a reference plan or clean frame as approved.
- Add
must_not_animate rules for any board with labels, arrows, panel borders, floor plans, timing notes, or tables.
- Add exact-card model pack recipes for Sora, Seedance, Kling, Runway, and Veo. Keep Kling VIDEO 3.0 separate from its legacy 5/10-second card; keep Runway Gen-4.5 generation, Aleph 2.0 Web, Aleph 2.0 API, and deprecated Gen-4 Aleph separate; use stable Veo 3.1
*-001 cards instead of preview aliases.
- Layout spec must protect readability and video-model safety.
- Represent the reference pack as asset nodes with explicit graph edges: planning-only, reference-only, or direct video input.
- Preserve prompt optimizer provenance if an optimized prompt is produced: keep the JSON-first source intent and the optimized text output.
skill_run_receipt
Record reference pack decision, visual output mode, execution capability, exact capability cards, source tiers/dates/conflicts, rights status, asset output statuses, rejected layouts, storyboard/clean-frame separation, direct input policy, optional canvas graph receipt, user decision gate status, QA status, and next_recommended_skill: image-prompt-compiler.