@aswless_854771076/ai_short_studio_cli 0.1.98 → 0.1.100

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/generate-storyboard-images/prompts/panel_grid_image.zh.txt

@@ -0,0 +1,77 @@
+你是一名短剧分镜故事板视觉导演。请根据给定单镜头分镜 JSON 和原文片段，生成一张包含 {panel_grid_size} 个分镜格的宫格图。
+
+【⚠️ 参考图使用规则 - 最高优先级，必须严格遵守】
+本次生成会附带角色/场景参考图，所有分镜格都必须严格参照这些参考图绘制，禁止凭空想象人物或场景：
+- 角色参考图：每个分镜格中角色的外貌、面部特征、发型、服装、配饰、体型都必须与角色参考图严格一致，不得改动或重新设计。
+- 场景参考图：用于参考环境的构图布局、建筑结构、风格与氛围；默认锁定同一机位、轴线、景别和背景关系，场景的视觉特征必须来自参考图，不要直接把参考图整张贴在背景上。
+- 参考图上方的文字标签标注了角色/场景名称，请与分镜 JSON 中出现的角色/场景对应；文字标签仅供识别，禁止画入图中。
+- 美术风格、线条、色彩、质感、光影都必须与参考图保持一致，禁止出现与参考图风格不符的画面。
+- 各分镜格只表现同一连续镜头内的时间推进和动作分解，人物与场景的视觉特征必须始终锁定参考图：跨所有分镜格保持同一人物、同一服装、同一场景，不允许某一格偏离参考图另起一套形象。
+
+【⚠️⚠️⚠️ 全局画风一致性 - 强制铁律】
+整张宫格图（包括所有分镜格中的主角、配角、路人、群众、背景、道具、环境元素）必须使用完全统一的美术风格：
+- 没有专属参考图的路人/配角/群众，必须严格套用主角参考图的画风（同一种描边、同一种上色、同一种渲染质感），严禁自行发挥成另一种画风
+- 所有分镜格都必须在同一画风下：主角是 2D 动漫，则所有格子里的所有人物、所有背景都必须是 2D；主角是写实风则全员写实
+- 严禁出现任何 3D/CG/CGI/Blender/皮克斯 3D 卡通/虚幻引擎/照片写实质感混入 2D 画面；严禁 2D 人物贴在 3D 背景上
+- 严禁角色手绘风而道具/背景是照片质感；严禁同一画面内出现多种渲染质感
+- 线条粗细、上色方式（平涂/渐变）、阴影软硬、高光处理必须全图统一
+- 反 AI 味：禁止过度锐化、过度 HDR、过饱和、塑料皮肤、蜡像感、油腻质感、多余手指、畸形五官
+- 光影方向和强度在同一分镜格内必须一致
+
+要求：
+1. 整张图为一张完整宫格图，不是单个镜头。
+2. 按 {grid_layout} 排列，共 {panel_grid_size} 个分镜格，阅读顺序从左到右、从上到下。
+3. 这 {panel_grid_size} 个分镜格是同一连续镜头的时间关键帧。默认保持相同机位、轴线、景别、镜头焦段和背景关系，只推进动作、表情、口型、走位及环境次级运动。只有 panel.camera_move 或 composition_reference 明确要求时，才按其轨迹连续改变构图；不得自行切成特写、插入镜头或其他视角。
+4. 全图保持统一画风（以已选画风合同与参考图为准：{style}）——所有分镜格、所有人物、所有背景元素执行同一 medium、rendering_method、surface_language 与 image_formation；参考图若来自旧画风或与当前合同冲突，必须停止生成并退回资产选版，不得静默跟随旧图。
+5. 整张图片比例为 {aspect_ratio}（每个子格被压缩到对应行列大小）。
+6. 不要生成无关中文或英文文字、水印、编号、对白气泡和界面元素；当 panel 明确要求系统文字或界面内容时，必须逐字保留并按 grid_plan 的时间顺序完整显示，不得截断括号、引号或数字。
+7. 分镜格之间需要有清晰边界，但不要让边框压过主体。
+8. 角色外貌、服装、场景和道具需要在多个分镜格中保持一致，且必须以参考图为准。
+9. 分镜 JSON 的 context.neighbor_panel_continuity 仅用于保持镜头语言连贯，禁止把相邻镜头剧情内容画进本宫格图。
+10. 如果「原文片段」包含台词内容，请根据分镜格的时间顺序，在对应格子中表现角色说话的口型和表情变化：
+    - 开始说话：嘴巴微张，表情起始状态
+    - 说话中：嘴巴开合变化，表情随情绪递进
+    - 说完：嘴巴闭合或保持最终表情
+    确保角色口型变化与台词内容的情感和节奏匹配。
+11. 当 panel 或原文明确“共 N 个座具、其中 X 个已坐、Y 个空置”时，每格必须精确画出 N 个座具：X 名指定人物的骨盆和腿部真实落在各自座具上，人物前方不得再出现对应座具的空置副本；只允许其余 Y 个座具无人占用。人数、座具数、占用数和空置数任一不符都视为失败。
+12. 严格执行角色引用中的 `performance_constraints`。心理变化按“刺激/信息—眼神转移—表情面具—短暂泄露—决定/动作—残留”分配到关键格；眼神落到具名人物、道具或空间目标，保持画面方向正确，非主观镜头或直视观众不得看镜头。近景/特写一次只画一至两个面部区域的轻微变化，中远景改用头部、呼吸姿态、停顿、距离和走位；听者保持独立注意和反应，不复制说话者动作。只有档案触发条件成立时才使用 habitual_actions，始终避开 forbidden_performance。
+
+【构图参考 - 最高优先级】
+- 若分镜数据包含 composition_reference，必须严格遵循其机位构图和角色站位。
+- composition_reference.primary_camera 为主机位，按其 FOV、机位坐标、目标点及角色站位/朝向构图。
+- composition_reference.alternate_views 列出同镜头其他机位参考图（含特写、反打等）；参考图按该数组顺序提供，需保持角色形象、光影和风格一致，整体构图以主机位为准。
+- 坐标系单位为米，y 轴向上，z 轴负方向为镜头前方，x 轴向右。
+
+【风格与场景冲突处理】
+- {style} 只作为画风、线条、色彩、光影和质感参考；如果其中包含与分镜地点冲突的具体场景元素，必须以分镜 JSON 的 panel.description/location 为准。
+- visual_bible 独立控制色板、影调、光线连续性和运动语言；不得把它改写成新地点、新角色或另一种绘画媒介。
+- 例如分镜是仙侠宫殿、室内、山林或街巷时，不要因为风格词里出现其他地点名而改画成无关场景。
+
+精简分镜 JSON：
+{storyboard_text_json_input}
+
+【宫格关键帧执行】
+- 分镜 JSON 中的 grid_plan.cells 已按 cellIndex 给出每个格子的时间顺序和关键帧意图，请按顺序绘制。
+- 第 1 格是起始瞬间，最后 1 格是收束瞬间，中间格用于动作、表情、口型、走位和机位的连续过渡。
+- 每格都必须服务于同一个连续镜头，不要画成多个互不相关的故事板镜头。
+
+【空间账本 - 出图前内部建立，不在画面中写字】
+- 为每个具名人物锁定：唯一身份、画面左/中/右、前/中/后景、身体朝向、脸朝向、视线目标、双脚落点、入出画边；只有 grid_plan 明确指定的移动者可以改变位置，其他人跨格固定。
+- 为每个关键道具锁定：唯一数量、初始位置、所有者、哪只手接触、接触前/中/后状态和镜尾位置。禁止复制、消失、换手、悬空或在人物前方再生成一个空置副本。
+- 多人镜先锁定总人数与从左到右顺序；座次镜同时锁定座具总数、占用者、空席编号和每人的坐姿接触。任何格都不得把同一人物画两次、把两人融合、交换身份或增添无名人物。
+- 屏幕方向以最终画面为准，不把世界坐标左右直接当成画面左右；先用 composition_reference 的相机映射站位，再保持轴线、视线和运动方向。若输入不足以唯一确定这些关系，停止生成并把缺项退回 unresolved，不得猜测。
+- 剧情特效只有在原文或已选道具/场景资产明确存在时才出现，并锁定发射源、路径、目标、数量、颜色与结束状态；禁止把分镜边框、箭头、圈线、色块或导演标注解释成门形光效、能量、烟雾或实体道具。
+
+【⚠️ 构图契约的时间化编译 - 格间几何变化】
+- 若分镜 JSON 的摄影规则含 composition_contract，或专业分镜在 photographyPlan.composition、visual_plan.composition 中给出构图合同，它是本镜已确认构图合同。若仅有 plan 期瘦合同 visual_plan.composition（只含 primary/secondary/reason/evidence），不要求 11 个几何字段齐全：缺失几何从 panel 与原文推断，推断不出的细节按旧分镜回退处理，不编造字段值。primary/secondary 构图 code 与构图模式名不得作为文字出现在任何格子里，也禁止只交「电影感」「压迫」「孤独」等抽象词；所有格子必须把合同翻译为可见几何并沿时间推进：
+  - 第 1 格为首帧构图：read_order 的第一眼落点、每个角色的 subject_placement（左右/前后景/仰平俯）、eyeline_room 视线与留白、depth_layers 前中后景内容与虚实、light_and_color_geometry 主光方向与剧内来源一次到位；
+  - 中间格沿 action_vector 连续变化：逼近/远离、入画/出画边、前后景穿插、遮挡何时出现及何时让出主体都必须可见；动作镜把蓄力—接触—结果—反应分配到相邻格，接触点与受力结果各自有清晰格子，过程中第一眼信息始终可读；
+  - 最后一格落到 tail_state：人物位置、朝向与剩余张力短暂稳定；
+  - 光源方向、视线轴线、屏幕方向与角色左右关系跨所有格子一致；action_vector 为 null 时只允许合同与分镜允许的次级运动，禁止自造推镜、环绕、漂浮或额外动势。
+- 证据门为 OR，对每一格生效：evidence 中 composition-only/set_dressing/weather_atmosphere/action_blocking 各等级必须在原文明确描述或所附已选角色/场景/道具资产中存在，二者任一即可（不要求同时具备，对齐 director-composition.md §0.2）；action_blocking 等级还可由已确认调度事实支撑，已确认调度事实不能替代天气/陈设所需的原文或资产证据；复合证据以 + 连接时逐组件成立。雨、雪、雾、烟尘、霓虹、镜面、门窗廊柱、武器等 weather_atmosphere/set_dressing 元素只能凭原文或已选资产入画，群体等 action_blocking 元素还可由已确认调度事实支撑；身份一致性与场景结构所要求的既有参考资产规则不受影响。任何一格缺证据都不得用想象补齐：停止生成，把缺失证据作为 unresolved 退回分镜阶段补证据或更换构图。禁止模型自行补雨、雪、雾、霓虹、镜面、烟尘、门窗廊柱、武器、群体或多余人物。
+- ai_risks 逐格可见防控：前景中的遮挡物不超过一侧 1/3 屏宽或仅占顶部一条、不盖脸也不盖关键反应，留白区不填物、接触点清晰、动势边缘留 10%、相邻格屏幕方向一致。
+- 竖屏 9:16：关系格用前后纵深、越肩或上下站位，不硬挤左右；关键脸、证据道具、接触点放在每格中上安全区，双眼完整可读；群戏沿格子拆成群体建立—主角动作—反应/结果，不横向平铺多人。
+- 旧分镜无合同字段时按 panel 与 grid_plan 原文绘制，不编造 code 或构图模式名。
+
+原文片段：
+{source_text}

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/generate-storyboard-images/prompts/panel_storyboard_image.en.txt

@@ -0,0 +1,109 @@
+You are a professional film storyboard director. Using the single-shot panel JSON, source text, reference images, and professional cinema knowledge, create one production-ready storyboard sheet driven by narrative beats, with exactly {panel_grid_size} frames. Do not sample frames at fixed time intervals.
+
+[EXECUTION PRIORITY]
+1. The primary camera, blocking, facing, and spatial axis in `composition_reference`.
+2. Character and location reference-image consistency.
+3. Story facts and reading order in `panel`, `source_text`, and `grid_plan`.
+4. Professional guidance, output constraints, terminology, and prohibitions in `cinemaKnowledge`.
+5. The style directive {style}. Lower-priority instructions must never override higher-priority constraints.
+
+[CANONICAL SHOT-METHOD ENUMS]
+{cinematography_vocabulary}
+- Select shot scale, camera angle, and camera movement from the enums above according to narrative need. Never distribute them evenly or stack techniques for spectacle.
+- Describe `shot_type` as “camera angle + shot scale” and use exactly one canonical movement name for `camera_move`; use “locked-off” when the camera does not move.
+
+[SCOPE]
+- The input describes one existing storyboard shot. Every frame must serve that shot; never expand it into unrelated scenes or a different plot event.
+- Treat `panel`, `source_text`, and `grid_plan.cells[].keyframe` as the sole narrative backbone and render them in ascending `cellIndex` order. Do not add, remove, reorder, or alter story events; only translate text into visible action, expression, composition, and relationship changes.
+- When the story fixes the total, occupied, and empty seat/support counts, treat occupancy as a physical hard constraint: each specified character's pelvis and legs must rest on their own support, no empty duplicate support may appear in front of them, and total/occupancy state must remain exact in every cell.
+- Arrange exactly {panel_grid_size} frames as {grid_layout}, reading left-to-right and top-to-bottom. The user selected this count; never add, remove, leave blank, or duplicate frames.
+- Each frame represents an independently valuable narrative beat or key pose, not an evenly spaced sample from a frame rate.
+- Clearly distinguish a cut from continuous camera movement. Unless the input explicitly requires a camera-position change, preserve continuous action and spatial relations within the same shot. If a continuous move reveals no new information, do not split it into near-duplicate frames.
+
+[NARRATIVE-BEAT PLANNING]
+Plan the beats internally before drawing; do not output the analysis:
+1. Extract cause, action, discovery, reaction, emphasis, relationship change, and resolution from `panel.description`, `panel.source_text`, character action, dialogue, props, and environmental changes.
+2. Map beats to the `cellIndex` order in `grid_plan.cells`. A `keyframe` is only a coarse boundary cue; turn it into a visible, performable story change using the actual plot.
+3. Every frame must change at least one of: shot scale, camera position, observed subject, key action pose, expression, character relationship, or narrative information.
+4. Never use nearly identical composition in three consecutive frames. Never duplicate a standing pose and change only arrows or text.
+5. If there are fewer natural beats than frames, subdivide only meaningful progressions such as “starting pose—turning point—result” or “looks—what is seen—reaction.” Never invent micro-movements with no information gain.
+6. The board must be so economical that removing any frame would lose narrative information. Otherwise merge that frame with its neighbor and use the released frame for a real information change.
+
+Use establishment, trigger, POV, reaction, emphasis, relationship, and resolution beats only when the story needs them; do not fill a checklist:
+- Establishment: characters, environment, entrance, lighting, and spatial relations.
+- Trigger: a sound, person, action, or environmental change.
+- POV: clearly show what the character actually sees.
+- Reaction: a visible, specific response from the key character.
+- Emphasis: reveal a new person, important prop, threat, or identity cue.
+- Relationship: place principal characters in one composition to create a new relation or conflict.
+- Resolution: complete the action, emotional landing, or strongest visual memory.
+
+[SHOT AND PERFORMANCE DESIGN]
+- Never rewrite dialogue or draw it as text. Express the original delivery only through the matching character's mouth shape, expression, eyeline, and body language; never invent speaking action when no dialogue exists.
+- Shot scale must serve the story and may vary among extreme wide, wide, medium, medium close-up, close-up, and insert, but never vary without purpose.
+- Use at most one primary camera movement from the enum above per frame. Never annotate contradictory camera moves.
+- Use tracking or a long take only when a moving subject needs continuous presentation.
+- Show a head turn through its starting pose, decisive turn, and final eyeline; do not fragment it into many tiny changes.
+- A significant character's first appearance needs a clear subject frame. Do not keep other characters looking off-screen without revealing what they see.
+- Establish focus through scale, eyeline, occlusion, depth, and light rather than explanatory text.
+- Enforce each character reference's `performance_constraints`. Gaze lands on a named person, prop, or spatial target; across frames show only the meaningful hold, glance, avert, or return states with correct screen direction. Never look into the lens without explicit POV or direct-address evidence.
+- Distribute psychological acting as stimulus/information → gaze shift → social mask → brief leak → decision/action → residue. A close-shot micro-expression changes only one or two facial regions at low amplitude rather than exaggerating the whole face; at wider scales use head, breath, posture, pause, distance, and blocking. A listener keeps independent attention and reaction instead of copying the speaker.
+
+[SPATIAL AND VISUAL CONTINUITY]
+- Before drawing, establish every character's blocking, facing, entrance, left-right relationship, travel direction, and eyeline from the JSON.
+- Preserve the 180-degree axis unless the input or `cinemaKnowledge` explicitly calls for crossing it; any crossing needs a readable transition.
+- Face, hair, headwear, costume, body type, height, and palette must match references and remain stable across all frames.
+- Keep attendant count, order, costumes, and key props stable.
+- Preserve architecture, stairs, doors, railings, entrances, and light direction. Keep only environment elements that serve the story.
+- `context.neighbor_panel_continuity` is only for shot continuity; never draw adjacent-shot plot content.
+
+[SPATIAL LEDGER — BUILD INTERNALLY BEFORE RENDERING; DO NOT WRITE IT INTO THE IMAGE]
+- Lock each named character to one identity, screen left/center/right, foreground/midground/background, body facing, face direction, eyeline target, foot position, and entry/exit edge. Only movers explicitly named by `grid_plan` may change position; everyone else remains anchored across frames.
+- Lock each key prop to one count, initial position, owner, contacting hand, before/contact/after state, and tail position. Never duplicate, vanish, swap hands, float, or draw an empty duplicate support in front of its occupant.
+- In group shots lock exact headcount and left-to-right order first. For seating, also lock total supports, occupants, empty-seat identities, and physical pelvis/leg contact. Never duplicate one character, merge two people, swap identities, or add an unnamed person.
+- Screen direction means final-image direction, not raw world-coordinate direction. Map blocking through the `composition_reference` camera before locking axis, eyelines, and travel. If the inputs cannot determine these relationships uniquely, stop and return the missing facts as unresolved instead of guessing.
+- A story effect appears only when source text or a selected asset explicitly supports it, with a locked source, path, target, count, color, and end state. Never interpret borders, arrows, circles, color marks, or direction notes as portal light, energy, smoke, or a physical prop.
+
+[COMPOSITION CONTRACT DELIVERED FRAME BY FRAME]
+- If the panel JSON carries composition_contract in its photography rules, or a professional storyboard nests a composition contract under photographyPlan.composition / visual_plan.composition, it is the confirmed contract for this shot. If only the thin plan-time visual_plan.composition exists (primary/secondary/reason/evidence), do not demand the 11 geometry fields: derive the missing geometry from the panel and source text, and where it cannot be derived treat the shot as a legacy fallback; never invent field values. The primary/secondary codes and pattern names must never enter the information strip or the colored production notes (the strip shows only shot number | shot scale | camera position | camera move; notes describe concrete geometry only), and bare abstractions ("cinematic", "oppressive", "lonely") are not instructions. Every main frame must translate the contract into visible geometry.
+- Beat mapping: an establishment frame states the opening composition and spatial relations — the read_order first-glance landing, subject_placement, eyeline_room, depth_layers, and light_and_color_geometry all at once; action frames split into anticipation—contact—result—reaction, with dedicated readable frames for the contact point and force result while the first-read information stays clear; the resolution frame lands on tail_state with position, facing, and residual tension briefly stable. When the sheet contains a cut, screen direction, entry/exit edges, and axis must stay consistent on both sides; crossing the axis requires support from the composition contract or cinemaKnowledge plus a readable transition.
+- Paint occluders and coverage, clean negative space, key-light direction with its in-world source, and foreground/midground/background blur per contract in every frame; frames whose action_vector is null add no motion energy, and shots with a static contract state locked-off camera — never self-invented push-ins, orbits, or drift.
+- The evidence gate is OR and applies per frame: composition-only/set_dressing/weather_atmosphere/action_blocking in evidence must be either explicitly described in the source text or present in the attached selected character/location/prop assets — either one suffices (they need not both be present, aligned with director-composition.md §0.2); the action_blocking level may additionally be supported by confirmed blocking facts, which cannot replace the source-text or asset evidence required for weather/set dressing; each component holds when compound evidence is joined with +. Rain, snow, fog, smoke/dust, neon, mirrors, doors/windows/columns, and weapons at the weather_atmosphere/set_dressing level may enter a frame only with source-text or selected-asset support, while action_blocking elements such as crowds may additionally rely on confirmed blocking facts; the existing reference-asset requirements for identity consistency and set structure remain unaffected. If any frame lacks evidence, stop producing the board and return the shot to the storyboard stage as unresolved with the missing evidence named, so evidence can be added or the pattern changed. Never add rain, snow, fog, neon, mirrors, smoke, doors/windows/columns, weapons, crowds, or extra people on your own.
+- Every ai_risks code needs a visible safeguard in each frame: foreground occluders cover at most one third of one side or only a thin top strip and never cover the face or key reactions, negative space stays empty, contact points read clearly, keep a 10% margin at motion edges, and keep screen direction consistent across frames.
+- Vertical 9:16: stage relationships through front/back depth, over-shoulder framing, or vertical placement instead of squeezing side by side; put key faces, evidence props, and contact points in each frame's upper-middle safe area with both eyes readable; split crowd scenes into beat frames such as crowd establishment, protagonist action, and reaction/result rather than lining up many figures horizontally.
+- Legacy panels without the contract follow the panel and grid_plan; never invent codes or pattern names.
+
+[INFORMATION STRIP]
+- Under each frame show only “shot number | shot scale | camera position | camera move”. Use `cellIndex` as the number and keep every label consistent with the drawing.
+- Never show timecodes, start/end times, or total duration in the storyboard image. Duration is controlled separately by the downstream video-generation step.
+
+[STORYBOARD LAYOUT AND DELIVERY-STYLE IMAGERY]
+- Use clean white or warm-white production-board paper, regular black rectangular borders, and even gutters. Overall aspect ratio is {aspect_ratio}; every frame has the same aspect ratio.
+- Reserve roughly 80% of each cell for the drawing and 20% for the information strip and production notes. Keep borders, gutters, and strips consistent.
+- Render every main frame directly in the project's target delivery medium and art treatment {style}. Execute its medium, rendering_method, surface_language, image_formation, and forbidden_substitutions. Never interpret “storyboard work sheet” as an automatic request for monochrome sketch, concept art, or live-action realism, and never let a template example override the user's selected medium.
+- Unless the approved delivery style explicitly calls for hand drawing, never place monochrome pencil, charcoal, grayscale hatching, paper texture, or draft treatment inside the main frames. Forbid comic dialogue bubbles, comic speed lines, poster layouts, and character-design sheets.
+
+[VIDEO-MODEL-SAFE LAYOUT]
+- This board will be used directly as a video-model reference. Forbid colored arrows, circles, path lines, color blocks, production notes, speed lines, and emotion symbols both inside the frames and along their edges. Express direction, eyeline, camera motion, and emotion through pose, screen position, depth, and lighting alone.
+- Use neutral black/gray text only in the information strip; do not color-code it red, green, blue, orange, or purple. Outside that strip, generate no explanatory text, dialogue bubbles, watermarks, headers, footers, or UI.
+- If an annotated human-review board is needed, save it as a separate review asset and never bind it as the `shot_board` in video-prompts.
+
+[REFERENCES AND PROFESSIONAL KNOWLEDGE]
+- Character identity, face, hair, costume, body type, location structure, and key props must strictly match attached references. Reference labels are identifiers only and must not be copied into the board.
+- If the JSON contains `cinemaKnowledge`, obey `guidance` and `outputConstraints`, translate `promptFragments` into concrete cinematography, and strictly avoid `negativeConstraints`. Never change story facts or character identity.
+- Use {style} to define the target delivery medium, material rendering, and art treatment inside every main frame. Borders, information strips, and sparse production notes remain planning-only structure.
+- `visual_bible` independently constrains the main frames' palette, tone, lighting continuity, and motion intent; it must not change the story location or character identity.
+- If `composition_reference` exists, obey its primary-camera FOV, position, target, blocking, and facing. Alternate views are continuity references only.
+
+[FINAL SELF-CHECK]
+1. The board is economical enough that removing any frame would lose narrative information.
+2. Adjacent frames have a clear action, viewpoint, subject, relationship, or information change.
+3. Viewers can understand what characters see and what happens without reading notes.
+4. Information strips contain no timecodes; shot scale, camera position, and camera move match the images.
+5. Characters, props, architecture, lighting, and screen direction remain continuous.
+
+Compact panel JSON:
+{storyboard_text_json_input}
+
+Source text:
+{source_text}

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/generate-storyboard-images/prompts/panel_storyboard_image.zh.txt

@@ -0,0 +1,109 @@
+你是一名专业影视分镜导演。请根据单镜头分镜 JSON、原文片段、参考图与专业电影知识，生成一张包含 {panel_grid_size} 格、可直接用于现场沟通的“叙事节拍驱动”故事板工作稿。不要按固定秒数机械抽帧。
+
+【执行优先级】
+1. composition_reference 的 primary_camera、角色站位、朝向与空间轴线。
+2. 角色和场景参考图的一致性。
+3. panel、source_text 与 grid_plan 的剧情事实和阅读顺序。
+4. cinemaKnowledge 的专业指导、输出约束、专业表达和禁止项。
+5. 风格指令 {style}。低优先级内容不得覆盖高优先级约束。
+
+【规范镜头方式枚举】
+{cinematography_vocabulary}
+- 从以上枚举中按叙事需要选择景别、视角和运镜，不要平均分配或为了炫技而堆叠。
+- 每格使用“视角+景别”描述 shot_type，camera_move 只使用一个规范运镜名称；无机位运动时使用“固定”。
+
+【适用边界】
+- 当前输入描述一个已有分镜镜头；所有格子共同服务这个镜头，不得扩写成无关场景或另一段剧情。
+- 以 panel、source_text 和 grid_plan.cells[].keyframe 为唯一叙事骨架，按 cellIndex 升序呈现；不得新增、删除、交换或改写剧情事件，只能把文字转成可见的动作、表情、构图和关系变化。
+- 当剧情明确座具总数、已坐数与空置数时，把占用关系当作物理硬约束：指定人物的骨盆和腿部必须落在各自座具上，人物前方不得出现对应座具的空置副本；总数与占用状态必须逐格精确一致。
+- 严格按 {grid_layout} 输出 {panel_grid_size} 格，阅读顺序从左到右、从上到下。格数由用户配置决定，不得增减、留空或重复填格。
+- 每格对应一个具有独立叙事价值的关键节拍或关键姿态，不是按帧率抽取的等间隔画面。
+- “切镜”与“连续运镜”必须明确区分。除非输入明确要求切换机位，否则保持同一镜头内的连续动作与空间关系；连续运镜没有产生新信息时，不得拆成多个近似格子。
+
+【叙事节拍规划】
+绘制前先在内部完成节拍拆解，不要输出分析过程：
+1. 从 panel.description、panel.source_text、角色动作、台词、道具和环境变化中提取起因、动作、发现、反应、强调、关系变化与收束。
+2. 将节拍映射到 grid_plan.cells 的 cellIndex 顺序；keyframe 只是粗粒度边界提示，必须结合剧情改写成可见、可执行的画面变化。
+3. 每格至少改变一项：景别、机位、观察主体、动作关键姿态、表情、角色关系或叙事信息。
+4. 禁止连续三个格子使用几乎相同的构图；禁止只复制人物站姿再修改箭头或文字。
+5. 若自然节拍少于格数，只能拆分有意义的“起始姿态—关键转折—动作结果”“看到—所见—反应”等子节拍，不得制造无信息增量的微小动作。
+6. 删除任意一个格子都会损失叙事信息；否则该格应与相邻格合并，并把腾出的格子用于真正的信息变化。
+
+可按剧情需要选择建立、触发、主观、反应、强调、关系和收束节拍，不要机械套齐：
+- 建立：交代人物、环境、入口、光线与空间关系。
+- 触发：声音、人物、动作或环境发生变化。
+- 主观：明确展示角色究竟看到了什么。
+- 反应：展示关键角色可见且明确的反应。
+- 强调：突出新人物、重要道具、威胁或身份信息。
+- 关系：让主要人物进入同一构图，形成新的关系或冲突。
+- 收束：完成动作、情绪落点或视觉记忆点。
+
+【镜头与表演设计】
+- 台词不得改写或画成文字；只通过对应角色的口型、表情、视线和肢体动作表现原文语气，没有台词时不得虚构说话动作。
+- 景别必须服务叙事，可在远景、全景、中景、中近景、近景和特写之间变化，但不得为变化而变化。
+- 每格最多使用一种上方枚举中的主要运镜；禁止同时标注互相矛盾的运镜。
+- 只有运动主体需要连续展示时才使用跟拍或长镜头。
+- 角色转头只表现起始姿态、关键转折和最终视线，不拆成大量微小变化。
+- 重要人物首次出现时必须有清晰主体画面；不能只让其他角色持续看向画外而不展示所见对象。
+- 用景别、视线、遮挡、前后景和明暗建立视觉焦点，不依赖文字补救含糊构图。
+- 严格执行角色引用中的 `performance_constraints`：眼神必须落到具名人物、道具或空间目标，跨格显示持续注视、扫视、回避或返回的关键状态，并保持画面方向正确；没有主观镜头或直视观众依据时不得看镜头。
+- 心理戏按“刺激/信息—眼神转移—表情面具—短暂泄露—决定/动作—残留”分配到有叙事价值的格子。近景/特写的微表情一次只画一至两个区域的轻微变化，不用夸张整脸表情；中远景用头部、呼吸姿态、停顿、距离和走位表达。听者必须有独立注意和反应，不复制说话者动作。
+
+【空间与视觉连续性】
+- 开始前根据 JSON 确定所有角色的站位、朝向、入口位置、人物左右关系、行进方向和视线方向。
+- 遵守 180 度轴线，除非输入或 cinemaKnowledge 明确要求越轴；越轴时必须设计可理解的过渡。
+- 同一角色的脸型、发型、发饰、服装、体型、身高和配色必须与参考图一致，并在全部格子中保持稳定。
+- 随从人数、排列顺序、服装、关键道具不得随机变化。
+- 建筑结构、台阶、门、栏杆、入口和光源方向必须连续；环境只保留对剧情有用的元素。
+- context.neighbor_panel_continuity 仅用于镜头衔接，禁止画入相邻镜头的剧情。
+
+【空间账本 - 出图前内部建立，不在画面中写字】
+- 为每个具名人物锁定唯一身份、画面左/中/右、前/中/后景、身体朝向、脸朝向、视线目标、双脚落点和入出画边；只有 grid_plan 明确指定的移动者可以改变位置，其他人跨格固定。
+- 为每个关键道具锁定唯一数量、初始位置、所有者、接触手、接触前/中/后状态和镜尾位置；禁止复制、消失、换手、悬空或给已占座人物再画一个空置副本。
+- 多人镜先锁定总人数与从左到右顺序；座次镜同时锁定座具总数、占用者、空席编号及骨盆和腿部接触。任何格都不得把同一人物画两次、融合两人、交换身份或增添无名人物。
+- 屏幕方向以最终画面为准，不把世界坐标左右直接当成画面左右；先用 composition_reference 的相机映射站位，再保持轴线、视线和运动方向。输入不足以唯一确定时停止出图并退回 unresolved，不得猜测。
+- 剧情特效必须有原文或已选资产证据，并锁定发射源、路径、目标、数量、颜色与结束状态；禁止把分镜边框、箭头、圈线、色块或导演标注解释成门形光效、能量、烟雾或实体道具。
+
+【构图契约逐格兑现】
+- 若分镜 JSON 的摄影规则含 composition_contract，或专业分镜在 photographyPlan.composition、visual_plan.composition 中给出构图合同，它是本镜已确认构图合同。若仅有 plan 期瘦合同 visual_plan.composition（只含 primary/secondary/reason/evidence），不要求 11 个几何字段齐全：缺失几何从 panel 与原文推断，推断不出的细节按旧分镜回退处理，不编造字段值。primary/secondary 的 code 与构图模式名不得进入信息条或彩色导演标注（信息条只写镜号｜景别｜机位｜运镜，标注只写具体几何动作），也禁止只交「电影感」「压迫」「孤独」等抽象词；每格主画面必须把合同翻译为可见几何。
+- 节拍映射：建立格交代首帧构图与空间关系——read_order 第一眼落点、subject_placement、eyeline_room、depth_layers、light_and_color_geometry 一次到位；动作格按蓄力—接触—结果—反应拆分，接触点与受力结果各自成格且第一眼信息始终可读；收束格落到 tail_state，人物位置、朝向与剩余张力短暂稳定。板内含切镜时，切点两侧保持屏幕方向、出入画边与轴线一致；越轴必须有构图合同或 cinemaKnowledge 依据并给出可读过渡。
+- 每格的遮挡物与遮挡比例、留白纯净度、主光方向与剧内来源、前中后景虚实按合同执行；action_vector 为 null 的格子不制造动势，静止合同的镜头写明固定机位，禁止自行缓推、环绕或漂浮。
+- 证据门为 OR，逐格生效：evidence 中 composition-only/set_dressing/weather_atmosphere/action_blocking 必须在原文明确描述或所附已选角色/场景/道具资产中存在，二者任一即可（不要求同时具备，对齐 director-composition.md §0.2）；action_blocking 等级还可由已确认调度事实支撑，已确认调度事实不能替代天气/陈设所需的原文或资产证据；复合证据以 + 连接时逐组件成立。雨、雪、雾、烟尘、霓虹、镜面、门窗廊柱、武器等 weather_atmosphere/set_dressing 元素只能凭原文或已选资产入画，群体等 action_blocking 元素还可由已确认调度事实支撑；身份一致性与场景结构所要求的既有参考资产规则不受影响。任何一格缺证据即停止出图，把缺失证据作为 unresolved 退回分镜阶段补证据或更换构图。禁止自行补雨、雪、雾、霓虹、镜面、烟尘、门窗廊柱、武器、群体或多余人物。
+- ai_risks 每个风险码在每格都要有可见防控：前景中的遮挡物不超过一侧 1/3 屏宽或仅占顶部一条、不盖脸也不盖关键反应，留白区不填物、接触点清晰、动势边缘留 10%、跨格屏幕方向一致。
+- 竖屏 9:16：关系镜用前后纵深、越肩或上下站位，不硬挤左右；关键脸、证据道具、接触点放每格中上安全区，双眼完整可读；群戏拆为群体建立、主角动作、反应/结果等节拍格，不横向平铺多人。
+- 旧分镜无合同字段时按 panel 与 grid_plan 处理，不编造 code 或构图模式名。
+
+【信息条】
+- 每格底部只显示“镜号｜景别｜机位｜运镜”。镜号使用 cellIndex，其他内容必须与实际画面一致。
+- 不得在故事板图片中显示时间码、开始结束时间或总时长；视频时长由后续视频生成节点单独控制。
+
+【故事板版式与成片风格画面】
+- 使用干净的白色或暖白色故事板工作纸，规则黑色矩形分框和均匀留白；整张图比例为 {aspect_ratio}，每格保持相同宽高比。
+- 每格约 80% 为主画面、20% 为信息和导演标注区，所有边框、留白和信息区保持统一。
+- 主画面必须直接使用项目的成片目标媒介与美术质感 {style}。逐项执行其中的 medium、rendering_method、surface_language、image_formation 和 forbidden_substitutions；不得把“故事板工作稿”自动理解成黑白草图、概念原画或真人写实，也不得用模板示例覆盖用户选定媒介。
+- 除非项目成片风格本身明确为手绘，禁止黑白铅笔、炭笔、灰阶排线、纸张纹理或草图质感进入主画面。禁止漫画对白框、漫画速度线、海报排版和角色设定表。
+
+【视频模型安全版式】
+- 本板将直接作为视频模型参考，主画面和画格边缘都禁止彩色箭头、圈线、路径线、色块、导演批注、速度线或情绪符号；所有方向、视线、运镜和情绪必须由人物姿态、画面位置、景深和光线直接表达。
+- 信息条仅用中性黑灰文字，不使用红、绿、蓝、橙、紫色编码；除信息条外不生成任何说明文字、对白气泡、水印、页眉页脚或 UI。
+- 若需要带标注的人类审片板，必须另存独立 review 资产，禁止把它绑定为 video-prompts 的 shot_board。
+
+【参考图与专业知识】
+- 角色外貌、面部、发型、服装、体型、场景结构和关键道具必须严格遵循所附参考图。参考图标签只用于识别，不要照抄进画面。
+- 若 JSON 包含 cinemaKnowledge，必须执行 guidance 和 outputConstraints，将 promptFragments 转化为具体镜头语言，并严格避开 negativeConstraints；不得改变剧情事实或角色身份。
+- {style} 决定每格主画面的成片目标媒介、材质和美术质感；分框、信息条和少量导演标注仅是工作稿结构。
+- visual_bible 独立约束主画面的色板、影调、光线连续性与运动意图，不得改变剧情地点或角色身份。
+- 若包含 composition_reference，严格遵循 primary_camera 的 FOV、机位、目标点、人物站位与朝向；其他机位仅用于连续性参考。
+
+【输出前自检】
+1. 删除任意一个格子都会损失叙事信息。
+2. 相邻格存在清晰的动作、视角、主体、关系或信息变化。
+3. 观众无需阅读标注，也能理解人物看到了什么、发生了什么。
+4. 信息条不含时间码，景别、机位和运镜与画面一致。
+5. 所有角色、道具、建筑结构、光线和空间方向保持连续。
+
+精简分镜 JSON：
+{storyboard_text_json_input}
+
+原文片段：
+{source_text}

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/ideate-drama-from-market/prompts/market_inspiration.en.txt

@@ -0,0 +1,23 @@
+# Codex execution contract: market_inspiration
+
+This file is read by the owning Skill to constrain Codex input, reasoning, and output. It is not a system prompt sent to an external text model. Braced fields are named input slots.
+
+[SAVED MARKET REPORT]
+{market_report}
+
+[USER-SELECTED TOPICS OR SIGNALS]
+{selected_market_focus}
+
+[USER-CONFIRMED CREATIVE BOUNDARIES]
+{user_decisions}
+
+Return strict JSON whose only top-level keys are: schema_version, report_ref, signals, hypotheses, decision, guardrails.
+
+Rules:
+1. schema_version is always market-inspiration.v1. After the user selects hypotheses, run `project-store.mjs market-report-ref <project-root> <report ID> <selected hypothesis ID>...`. Use that command's JSON output verbatim as report_ref; it may contain only report_id, snapshot_ids, generated_at, filters, sha256, and selected_hypothesis_ids. Never calculate the digest yourself, guess its algorithm, or invent fields.
+2. signals[].fact contains only report-backed facts. Each signal requires confidence, at least one evidence item, and at least one limitation. Preserve low confidence as low.
+3. In every hypothesis, inference is a bounded inference; creative_transformation, opening_hook, serial_engine, differentiation, and production_fit are original creative hypotheses. Cite at least one derived_signal_ids and include premises, risks, and validation_questions.
+4. Do not copy, rewrite, or combine ranking-work titles, character relationships, concrete plots, distinctive dialogue, or story structures. Do not promise revenue, media-buying results, or success probability.
+5. decision.selected_hypothesis_ids must exactly match report_ref. Select only hypotheses explicitly chosen by the user. All four guardrails booleans must be true.
+
+No Markdown, explanation, screenplay body, or external-model parameters.

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/ideate-drama-from-market/prompts/market_inspiration.zh.txt

@@ -0,0 +1,23 @@
+# Codex 执行合同：market_inspiration
+
+本文件由所属 Skill 读取，用于约束 Codex 的输入、判断与输出；不是发送给外部文本模型的 system prompt。花括号字段是具名输入槽。
+
+【已保存市场报告】
+{market_report}
+
+【用户选择的题材或信号】
+{selected_market_focus}
+
+【用户确认的创作边界】
+{user_decisions}
+
+只返回严格 JSON，顶层必须且只能包含：schema_version、report_ref、signals、hypotheses、decision、guardrails。
+
+规则：
+1. schema_version 固定为 market-inspiration.v1。用户选定假设后，先调用 `project-store.mjs market-report-ref <项目目录> <report ID> <已选 hypothesis ID>...`；report_ref 必须逐字段原样使用该命令的 JSON 输出，只能包含 report_id、snapshot_ids、generated_at、filters、sha256、selected_hypothesis_ids。不得自行计算摘要、猜测算法或杜撰字段。
+2. signals[].fact 只写报告可回链的事实；每项必须包含 confidence、至少一条 evidence 和至少一条 limitations。低置信度必须原样继承为 low，不得升级。
+3. hypotheses[] 中 inference 是有限推断，creative_transformation、opening_hook、serial_engine、differentiation 和 production_fit 是原创假设。每项必须引用至少一个 derived_signal_ids，并写明 premises、risks 和 validation_questions。
+4. 不得复制、改写或拼接榜单作品标题、人物关系、具体情节、独特台词或故事结构；不得承诺收益、投放或成功概率。
+5. decision 的 selected_hypothesis_ids 必须与 report_ref 相同；只选择用户明确选定的假设。guardrails 的四个布尔字段必须全部为 true。
+
+禁止 Markdown、解释、剧本正文和外部模型参数。

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/outline-drama-series/prompts/episode_split.en.txt

@@ -0,0 +1,42 @@
+# Codex execution contract: episode_split
+
+The owning Skill reads this file to constrain Codex inputs, decisions, and output. It is not a system prompt sent to an external text model. Braced fields are named input slots assembled from selected local project assets.
+
+Execution role: short-drama episode planner. From the approved brief, story bible, and source boundaries, produce the complete episode outline.
+
+[INPUT]
+{CONTENT}
+
+Rules:
+1. Use the confirmed episode count. Plan around the target duration without damaging actions, dialogue, or hooks merely to hit an exact second, and never infer episode count from a fixed word quota. By default the final cut may vary naturally by about ±15%; larger variance is not an automatic failure, but requires a recorded narrative reason and renewed user confirmation.
+2. Every episode must add an action, revelation, relationship change, or cost. Never repeat the prior episode's emotion without progression.
+3. opening_connection establishes or continues the prior hook; episode_goal is an active objective; turning_point changes the situation; ending_hook creates a question or action the next episode must answer.
+4. Fulfill the story bible's conflict ladder, foreshadowing, and payoffs. Do not spend the final climax early or omit required setup.
+5. For adaptations, source_range records exact start and end anchors. Ranges must cover the source in order without gaps or overlap. Original projects may use null.
+6. required_assets lists only canonical names of characters, scenes, and key props actually needed in that episode. Do not add visual descriptions or model parameters.
+7. Episode keys are ep-001, ep-002, and so on; order starts at 1 and is consecutive.
+
+Return strict JSON only:
+{
+  "episodes": [
+    {
+      "key": "ep-001",
+      "order": 1,
+      "title": "Episode title",
+      "logline": "One-sentence cause and effect",
+      "opening_connection": "How this episode establishes or continues the prior hook",
+      "episode_goal": "The active objective",
+      "beats": ["Visible action or information beat"],
+      "turning_point": "Required change in the situation",
+      "payoff": "Promise fulfilled in this episode",
+      "ending_hook": "Next-episode hook",
+      "source_range": {"start_marker": "verbatim anchor", "end_marker": "verbatim anchor"},
+      "required_assets": {"characters": [], "scenes": [], "props": []},
+      "estimated_duration_seconds": 90
+    }
+  ],
+  "coverage_check": {"complete": true, "gaps": [], "overlaps": []},
+  "continuity_check": {"valid": true, "issues": []}
+}
+
+No Markdown, explanations, or extra fields. Before returning, verify episode count, keys, order, source coverage, hook continuity, foreshadowing/payoff, and total duration.

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/outline-drama-series/prompts/episode_split.zh.txt

@@ -0,0 +1,42 @@
+# Codex 执行合同：episode_split
+
+本文件由所属 Skill 读取，用于约束 Codex 的输入、判断与输出；不是发送给外部文本模型的 system prompt。花括号字段是从当前项目已选资产组装的具名输入槽。
+
+执行职责：短剧分集策划。根据已确认的 brief、故事圣经和来源边界，生成完整分集目录。
+
+【输入】
+{CONTENT}
+
+规则：
+1. 严格使用 brief 中确认的总集数；单集时长以目标值为中心规划，不按固定字数自行推算集数，也不为精确卡秒破坏动作、对白或钩子。默认实际成片可在目标值约 ±15% 内自然浮动；超出不自动判失败，但必须记录叙事原因并重新请用户确认。
+2. 每集必须带来新的行动、信息、关系变化或代价；不得仅重复上一集情绪。
+3. opening_connection 明确承接上一集；episode_goal 给出本集主动目标；turning_point 改变局势；ending_hook 产生下一集必须回答的问题或行动。
+4. 按故事圣经兑现冲突阶梯、伏笔和回收。不得提前消耗最终高潮，也不得遗漏结局所需铺垫。
+5. 改编项目用 source_range 记录本集覆盖的原文起止锚点；必须按顺序完整覆盖且不得重叠。原创项目可将 source_range 设为 null。
+6. required_assets 只列本集实际需要的人物、场景和关键道具规范名，不写视觉描述或模型参数。
+7. episode key 固定为 ep-001、ep-002……，order 从 1 连续递增。
+
+只返回严格 JSON：
+{
+  "episodes": [
+    {
+      "key": "ep-001",
+      "order": 1,
+      "title": "集名",
+      "logline": "本集一句话因果",
+      "opening_connection": "如何承接上一集或建立首集",
+      "episode_goal": "人物本集主动目标",
+      "beats": ["可见行动或信息节拍"],
+      "turning_point": "不可省略的局势变化",
+      "payoff": "本集兑现",
+      "ending_hook": "下一集钩子",
+      "source_range": {"start_marker": "逐字锚点", "end_marker": "逐字锚点"},
+      "required_assets": {"characters": [], "scenes": [], "props": []},
+      "estimated_duration_seconds": 90
+    }
+  ],
+  "coverage_check": {"complete": true, "gaps": [], "overlaps": []},
+  "continuity_check": {"valid": true, "issues": []}
+}
+
+禁止 Markdown、解释和额外字段。输出前核对集数、key、order、原文覆盖、钩子承接、伏笔回收和总时长。

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/plan-drama-assets/prompts/asset_analysis.en.txt

@@ -0,0 +1,55 @@
+# Codex execution contract: asset_analysis
+
+The owning Skill reads this file to constrain Codex inputs, decisions, and output. It is not a system prompt sent to an external text model. Braced fields are named input slots assembled from selected local project assets.
+
+Execution role: short-drama asset coordinator. Build the episode asset plan from the approved source analysis, creative brief, selected script, director book, and existing local asset library. Extract only assets that visibly appear or are explicitly used. Do not generate media or rewrite the story.
+
+Episode key: {episode_key}
+
+Source analysis:
+{source_analysis}
+
+Creative brief:
+{brief}
+
+Selected episode script:
+{episode_script}
+
+Selected director book:
+{director_book}
+
+Existing local asset summary:
+{existing_assets}
+
+Rules:
+1. Merge aliases under stable character identities. Persistent costume or age changes are versions; temporary expressions and actions are not.
+2. Separate independently generated macro locations from micro areas that need their own composition. Camera-angle changes within one space are not new locations.
+3. Keep only movable, irreplaceable, cross-shot, or plot-critical props. Do not create standalone assets for ordinary furniture, tableware, clothing, or location fixtures.
+4. Put continuity-critical age, costume, genuine/fake, intact/damaged, or day/night states in versions. Never infer unsupported facts.
+5. Reuse semantically identical keys from existing_assets. Put same-name conflicts, uncertain identity merges, and ambiguous versions in unresolved for user confirmation.
+6. Every asset needs evidence with source, locator, and a short quote sufficient to prove the asset or version.
+7. Set derived_from to a reused asset key or null. Set selected_version only when an approved local version already exists; otherwise use null.
+8. status must be planned, ready, or blocked. Use blocked for unresolved identity, merge, version, or scope decisions and explain them in unresolved.
+
+Return strict JSON only, with exactly these top-level fields:
+{
+  "episode_key": "{episode_key}",
+  "characters": [
+    {
+      "key": "stable ASCII key",
+      "type": "character",
+      "name": "character name",
+      "evidence": [{"source": "script|director-book|source-analysis", "locator": "scene or paragraph", "quote": "short evidence"}],
+      "visual_description": "supported stable visual facts only; mark unknowns explicitly",
+      "versions": [{"key": "version key", "label": "version name", "trigger": "active scope", "evidence": "version evidence"}],
+      "derived_from": null,
+      "status": "planned",
+      "selected_version": null
+    }
+  ],
+  "scenes": [],
+  "props": [],
+  "unresolved": [{"id": "question key", "question": "question requiring user confirmation", "affects": ["asset key"]}]
+}
+
+scenes and props use the same field contract with type fixed to scene and prop respectively. Use empty arrays when absent. Do not omit fields or add fields outside this contract.

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/plan-drama-assets/prompts/asset_analysis.zh.txt

@@ -0,0 +1,55 @@
+# Codex 执行合同：asset_analysis
+
+本文件由所属 Skill 读取，用于约束 Codex 的输入、判断与输出；不是发送给外部文本模型的 system prompt。花括号字段是从当前项目已选资产组装的具名输入槽。
+
+执行职责：短剧资产统筹。根据已确认的来源分析、创作简报、单集剧本、导演本和现有本地资产库，建立本集资产计划。只提取画面中实际出现或被明确使用的资产，不生成图片，不改写剧情。
+
+分集 key：{episode_key}
+
+来源分析：
+{source_analysis}
+
+创作简报：
+{brief}
+
+已选单集剧本：
+{episode_script}
+
+已选导演本：
+{director_book}
+
+现有本地资产摘要：
+{existing_assets}
+
+规则：
+1. 人物按稳定身份合并别名；持续造型变化作为版本，不把临时表情或动作建成版本。
+2. 场景区分可独立生成的 macro 空间和需要独立构图的 micro 区域；同一空间的机位变化不是新场景。
+3. 道具只保留可移动、不可替代、跨镜持续或承担明确剧情功能的实体；普通家具、餐具、服装和场景设施不单建资产。
+4. 年龄、服装、真假、完好/损坏、昼夜等会影响连续性的状态写入 versions；没有证据不得猜测。
+5. 优先复用 existing_assets 中语义相同的 key；同名异物或同人不同身份必须列入 unresolved，等待用户确认。
+6. evidence 至少一条，包含 source、locator、quote；quote 只取足以证明资产或版本的短片段。
+7. derived_from 填复用资产 key，没有则为 null；selected_version 仅在现有资产已有批准版本时填写，否则为 null。
+8. status 只能是 planned、ready、blocked。存在身份、合并、版本或范围歧义时用 blocked，并在 unresolved 说明。
+
+只返回严格 JSON，不要 Markdown。顶层字段必须且只能如下：
+{
+  "episode_key": "{episode_key}",
+  "characters": [
+    {
+      "key": "稳定 ASCII key",
+      "type": "character",
+      "name": "人物名",
+      "evidence": [{"source": "script|director-book|source-analysis", "locator": "场次或段落", "quote": "短证据"}],
+      "visual_description": "仅写已知稳定视觉事实，未知项明确标记",
+      "versions": [{"key": "版本 key", "label": "版本名", "trigger": "生效范围", "evidence": "版本证据"}],
+      "derived_from": null,
+      "status": "planned",
+      "selected_version": null
+    }
+  ],
+  "scenes": [],
+  "props": [],
+  "unresolved": [{"id": "问题 key", "question": "需要用户确认的问题", "affects": ["资产 key"]}]
+}
+
+scenes 与 props 使用同一字段合同，type 分别固定为 scene、prop。无对应资产时输出空数组；不得省略字段，不得输出合同外字段。