| name | cinematic-blockout-director |
| description | Execute an approved relationship, action, spectacle, suspense, surreal, or design-display shot through a mandatory proxy-blockout-to-final workflow with domain-isolated references, prompt disclosure, structural preflight, independent generation, user selection, tool-capability fallback, and final-image QA. Use after narrative-shot-director approval, for human or non-human proxy blockouts, environment grayboxes, selected-blockout finals, scale locking, prompt compilation, or image critique. Never let identity references control presentation, and never skip directly from approved text to a final image. |
Cinematic Blockout Director
Turn an approved shot design into reliable images through two gates:
- generate and inspect composition blockouts;
- wait for explicit blockout selection, then generate and inspect final candidates.
This skill executes director decisions. It does not quietly replace the dominant subject, frozen action state, camera origin, crop, information sacrifice, performance intent, or light intent of an approved plan. It also does not let placeholder blockout clay, colors, neutral light, or missing decoration become final-scene authority.
Match prompt disclosure to the user's language
- Default the canonical prompt and visible disclosure to Simplified Chinese; use another user-requested language consistently when the tool supports it. The prompt shown to the user must equal the submitted prompt, not a summary.
- If the tool requires another language, show a complete direct translation with identical order and no omissions, and label the actual submission language.
- Before each image call, disclose the current stage, authorized count, reference list and role of every reference, director locks, allowed information loss, and the complete submitted prompt in the user's language or a complete direct translation.
- Disclosure does not create a new stop when the user already explicitly authorized the current generation. If the user asked to review the prompt first or has not authorized generation, disclose and wait.
- If the image tool returns a revised prompt, show its complete direct translation in the user's language separately. If it does not, say so; never imply access to a revised prompt that was not returned.
- Read authority-and-disclosure.md before compiling any image prompt.
Gate the input
Accept either:
- an explicitly user-approved
DIRECTOR SHOT PLAN from narrative-shot-director; or
- a sufficiently staged user-authored brief that directly specifies the intended camera/composition and requests blockout or image generation.
If the input is only an unstaged story event, do not generate and do not invent a complete director plan inside this skill. Route the story-to-shot task to narrative-shot-director, which must return a draft and wait for approval.
Do not treat a draft plan as approved unless the user explicitly selects it, approves it, or instructs this skill to use it. An explicit plan approval authorizes Stage 1 only: generate one blockout by default or the user's stated blockout count. Never interpret approval as authorization for a final image. Never begin final generation before the user explicitly selects a generated, readable blockout, even when only one blockout exists.
Route the execution request
Choose one mode:
- Final image request: run the mandatory blockout-to-final workflow.
- Blockout-only request: generate and inspect blockouts, then stop.
- Selected-blockout final request: verify that the user-selected blockout can be read as an image, then run the final stage regardless of prior quality observations.
- Prompt-only execution request: compile the approved plan into a blockout or final prompt without generating.
- Critique request: diagnose the supplied image without generating unless explicitly requested.
For image requests, use the available image-generation tool and inspect the actual pixels. Tool completion is not acceptance.
Check tool capability before promising the workflow
Before the first image call, establish whether the available tool supports image references, multiple references, the chosen prompt language, requested aspect ratio, and readable output inspection. Read tool-capability-fallbacks.md.
If a required capability is absent, state the limitation and choose the smallest honest fallback: prompt-only delivery, text-only identity generation with user permission, reduced reference set, approximate aspect, or blockout-only completion. Never claim reference isolation or exact continuity when the tool cannot receive the needed inputs.
Preserve the handoff contract
Normalize an approved director plan into three domains:
- DIRECTOR LOCKS: shot function, dominant subject, decisive instant, narrative focus pocket, camera origin, frame/crop, spatial layers, performance/contact intent, light intent, and allowed information loss;
- WORLD FACT LOCKS: identity, number, anatomy, required object, chronology, continuity, height, and scale;
- EXECUTION FLEX: only unspecified focal-length micro-adjustment, depth strength, material finish, model wording, and technical cleanup.
Apply authority inside separate domains rather than one global reference ranking:
- current user instruction overrides the corresponding domain;
- approved
DIRECTOR LOCKS plus the selected blockout control presentation: camera, frame, screen geography, pose, body/head axes, gaze, travel vector, crop, overlap, focus, light, and information loss;
- verified text facts control identity, count, anatomy, chronology, height, and scale;
- original character references control only visible identity, anatomy, costume, equipment, material, and design inside the approved crop;
- execution choices operate only inside unspecified implementation details.
Character references never control pose, body or head direction, gaze, travel direction, screen side, depth, camera, crop, completeness, sharpness, lighting, or visual weight. Do not describe a character reference as locking its pictured stance or facing direction.
Do not use character-reference visibility as a reason to widen, pull back, deepen focus, brighten suppressed regions, complete a cropped figure, or remove motivated occlusion. Invisible continuity is allowed; visible contradiction is not.
If two approved domains conflict or the geometry cannot physically satisfy them, stop before generation and return:
SHOT PLAN FEASIBILITY CONFLICT
Conflicting fields:
Why they cannot coexist:
Locks that remain unchanged:
Smallest option A:
Smallest option B:
STATUS: WAITING FOR USER DECISION
Never resolve such a conflict silently.
Enforce authorization and independence
- Default to one blockout and, after confirmation, one final.
- An unqualified request for
N images applies only to the current stage. Do not carry a blockout count into the final stage unless the user explicitly requests that final count.
- When the user authorizes
N > 1 blockouts, reserve the whole batch before the first call and complete all N independent calls regardless of any earlier visual deviation.
- During an authorized multi-blockout batch, never pause for confirmation, cancel later calls, or change the normalized prompt because of an earlier result.
- Present the whole batch with neutral observation notes after all calls finish. Do not assign
PASS/FAIL, rank eligibility, or prohibit selection.
- Generate each requested blockout in a separate image-generation call from the same normalized card and prompt. Do not attach, edit, derive from, or use another blockout candidate.
- Generate each requested final in a separate call from the selected blockout, original brief, complete final prompt, and original character references. Do not attach, edit, derive from, or use another final candidate.
- A failed final image never becomes an edit target or reference. Do not apply this rule to a user-selected blockout.
- In an authorized final-image batch, likewise inspect each result independently and continue remaining authorized calls after an individual failure. If all fail, stop.
- Outside an authorized batch, a failure ends the current image stage. Never retry automatically.
- Critique, prompt optimization, selection discussion, or
继续优化 does not authorize generation unless the wording clearly requests a new image.
Build the execution shot card
Read authority-and-disclosure.md and prompt-contract.md. Preserve the approved plan rather than reopening its director choices, then classify:
- FATAL LOCKS: identity, count, anatomy, decisive action/contact, required text, height/scale, key causal/spatial relationships, and explicit non-negotiables;
- REQUIRED-VISIBLE LOCKS: exact evidence that carries the beat;
- HEIGHT LOCKS: known heights and required relative stature;
- CONTINUITY LOCKS: wardrobe, equipment, color, and design facts that must be correct when visible but may recede;
- FINISH NOTES: grain, minor particles, surface density, and optical texture unless promoted by the user;
- EXECUTION CHOICES: only choices left unspecified by the approved plan or detailed user brief;
- FORBIDDEN INVENTIONS: additions that change the story or compete for attention.
Write every decisive visible action as a physical subject-body-object relation. Reject hovering hands, false contact, wrong object count, or evidence assigned outside the crop.
Before every image call, build a structured execution card and run scripts/validate_execution_card.py. A failed preflight stops generation. Never silence or manually waive a conflict; fix the card or ask the user.
Classify every important subject by proxy type before building the blockout: humanoid, quadruped, wheeled, flying, multi-legged, flexible-body, or mega-structure. Read subject-proxy-system.md. Do not force a non-human subject into a human mannequin.
Lock multi-character height
Read the height section in blockout-workflow.md whenever two or more important characters appear.
- Record stated heights exactly; use a relative tier rather than inventing unknown centimeters.
- Treat known height differences as fatal locks in blockout and final stages.
- Prove stature with at least two perspective-valid cues, using shared ground and fixed environmental scale.
- Do not substitute low angle, body width, foreground placement, or oversized armor for real height.
- Preserve stature through ground plane and vanishing geometry when figures occupy different depths.
- Preserve stature through torso/limb and environmental scale when a figure sits, kneels, crouches, bends, or is cropped.
Translate motion, focus, light, and finish
Use OFF, SLOW, FAST, or HIGH_SPEED. When layers differ, split SUBJECT_MOTION, RELATIONSHIP_OBJECT_MOTION, and BACKGROUND_MOTION. When any layer is not OFF, read motion-system.md. Never spread high-speed blur across a slower subject.
Use the approved camera, crop, spatial behavior, focus pocket, performance, and light intent as locks. Read cinematic-language.md only to implement unspecified optical, exposure, palette, and material behavior—not to replace an approved director decision.
When the user or approved plan names a visual focus, read focus-system.md. Convert it into a graded focus map around the smallest story-bearing feature or relationship, with one dominant read and at most two supporting reads.
Stage 1: generate blockouts
Read blockout-workflow.md and prompt-contract.md. Resolve the blockout card before invoking the image tool.
Preserve aspect, approved camera, perspective, scene masses, ground plane, important props, placement, pose, screen zone, depth, body axis, head axis, gaze, travel vector, contact, overlap, crop, and height locks.
Represent every human or humanoid as a smooth, solid-color, faceless anatomical muscle-volume mannequin:
- preserve only height, overall build, ribcage/pelvis volumes, limb proportions, joints, pose, and simplified contact-capable hands;
- use no armor, helmet, costume, fabric, hair, skin texture, identity silhouette, expression, facial features, eye sockets, mouth, or modeled face contour;
- do not use exposed-muscle anatomy texture;
muscle-volume means smooth body mass only;
- model a non-human appendage when it changes action, support, balance, contact, crop, or occlusion, using neutral jointed primitives rather than identity design;
- keep held story-critical objects as simple geometric proxies.
Represent non-human subjects with their proxy family from subject-proxy-system.md, preserving support points, articulation, turning or flight limits, footprint, mass distribution, contact surfaces, travel axis, and scale while suppressing identity design.
For spectacle, pursuit, disaster, crowd, or architecture-led shots, build an environment graybox with scale bands, terrain/architecture masses, paths, occluders, destruction or effect volumes, and density zones. A person-only white room is insufficient when the environment carries speed, danger, or scale.
Suppress final lighting, grade, particles, material finish, and decorative detail.
For multiple important characters, keep a stable solid map:
- primary: red;
- second: blue;
- third: green;
- fourth: yellow;
- additional important characters: magenta, cyan, orange, then purple;
- bystanders and unimportant people: white;
- environment and props: neutral white/gray.
Disclose the complete blockout prompt in the user's language or a complete direct translation, plus the reference list; pass structural preflight, then generate the authorized count with separate calls. Inspect each only to provide concise, non-binding observations from qa-rubric.md, report the color map, and stop for user selection. Never label a blockout passed or failed.
Selection gate
- Every generated, readable blockout remains selectable. Composition quality, wrong count, wrong height, wrong contact, or another visible deviation may be reported but never revokes the user's choice.
- The user is the sole blockout-selection authority. If the user says to use a previously criticized blockout, use it without requesting regeneration.
- Never begin final generation until the user explicitly selects or confirms one blockout.
- A selected blockout may be accompanied by textual corrections such as preserving its camera while correcting height or contact. Current user text overrides the conflicting part of the image.
- Block selection may be refused only when no image file exists, the file is unreadable, the image tool rejected it for safety, or there is no usable visual content to attach. Do not treat aesthetic or compositional disagreement as technical unreadability.
- If the user changes the blockout, generate a fresh candidate from the original approved plan, world facts, height locks, and current delta. Do not edit or attach the old blockout.
- Treat the selected blockout as composition/scale authority only, never identity, costume, material, lighting, or style authority.
Stage 2: generate finals
Before final generation:
- verify that the explicitly selected blockout is readable and record any current textual corrections without judging its eligibility;
- inspect identity references for every important character;
- request missing identity-critical references unless the user permits text-only generation;
- assign the blockout to composition/scale only, character references to identity/anatomy/costume, approved plan to presentation, and original brief to story facts and environmental meaning;
- build a
FINAL SCENE COMPLETION CARD from the original brief, world facts, and references before writing the final prompt;
- start a fresh generation rather than editing, repainting, or continuing the selected blockout;
- build and pass the structured authority/preflight card;
- disclose the complete final prompt in the user's language or a complete direct translation, plus the exact reference-role list;
- repeat the complete locks, completion card, and input roles on every independent final call.
The completion card must state:
- final environment type and required large architecture/furniture;
- production materials, surface age, wear, and lived-in density;
- set dressing and small semantic details allowed or required even when absent from the blockout;
- final motivated lighting, atmosphere, and palette;
- blockout properties forbidden from carrying over: clay/plastic material, character color map, neutral preview lighting, primitive geometry, blank surfaces, and under-furnished modeling-stage emptiness.
Use absence carefully:
- story-critical objects, large furniture, structural masses, and anything affecting occlusion or movement must be represented in the blockout;
- bedding, personal effects, panel seams, storage details, cables, wear, minor equipment, and similar non-blocking set dressing may be introduced in the final even when simplified or absent in the blockout;
- added details must not change the approved silhouette, action, path, crop, occlusion, height relation, or visual center.
The selected blockout is never a style, material, lighting, identity, or completeness reference. Do not tell the image tool to edit the blockout. Tell it to create a new final image while borrowing only the approved spatial construction.
Generate the authorized final count with separate calls and inspect every result using qa-rubric.md. A selected blockout does not excuse identity drift, incorrect height, false contact, director-lock drift, or technical failure.
Deliver or stop
- Blockout batch: show every generated candidate, color map, and concise non-binding observations, then ask which one the user wants to use.
- Passing final: lead with the image and briefly report preserved director/world locks, height evidence, motion evidence when applicable, and limitations.
- Failure inside a batch: record it, exclude it from later inputs, and continue all remaining authorized calls without changing the prompt; report the complete batch together.
- Single-final or all-final-candidate failure: stop, report failed gates and preserved strengths, propose a minimal textual delta, and wait for fresh authorization. This failure gate does not apply to blockout selection.
Follow applicable image-safety, privacy, likeness, copyright, and trademark rules. Do not expose lengthy internal analysis unless requested.