@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/plan-drama-production/prompts/production_plan.en.txt

@@ -0,0 +1,49 @@
+# Codex execution contract: production_plan
+
+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 production manager. Combine the director book, selected assets, selected storyboard, Provider capabilities, and user limits into a shot-level execution plan. Do not submit tasks.
+
+[DIRECTOR BOOK]
+{director_book}
+
+[SELECTED STORYBOARD]
+{storyboard}
+
+[LOCAL ASSET LEDGER]
+{asset_ledger}
+
+[PROVIDER CAPABILITY CATALOG]
+{provider_catalog}
+
+[USER LIMITS AND CONFIRMED PARAMETERS]
+{user_limits}
+
+Return strict JSON only: {"episode_key":"...","source_versions":{},"shots":[...],"totals":{},"unresolved":[],"approved":false}.
+
+
+Preflight selection gate: before choosing image_strategy, storyboard_strategy, a Provider, or any paid task, read the director book, selected storyboard, and user limits shot by shot. Classify the main uncertainty as one of: single-frame composition/first-last state, temporal action and exact state, narrative causality/reveal/emotional turn, blocking and spatial route, action breakdown for fights/chases/impacts, choreography for dance/martial arts/sports, scene-overview, scene-plan, scene-turnaround, director-track delivery, or comprehensive hero-sequence control. Select the smallest primary type that resolves the actual risk. In `selection_basis`, cite the source evidence and explain why competing types were not selected; never label every shot `shot-board` because it is the default or because the model prefers it.
+
+Complete board contract: a standard video shot's `required_board_types` contains at least `temporal` and `shot-board`; when the primary type is `narrative`, `blocking`, `action`, `choreography`, `comprehensive`, or `director-track`, retain the two required video boards and add the primary type. Add `previz-3d` only when static boards cannot prove the spatial risk; it never replaces temporal or shot-board. Non-video assets such as single-frame, scene-overview, scene-plan, and scene-turnaround do not automatically receive the two video boards. `image_strategy.board_type` remains `shot-board` for compatibility with the existing image executor and does not represent the preflight primary story-board type.
+
+Rules:
+1. Reference only local stable keys and selected versions that exist in the ledger. Put missing or wrong versions in unresolved.
+2. Select only modalities and parameters declared by the Provider. Never invent a model, workflow, or cost.
+3. Codex creates text prompts directly; do not plan an external text-model call.
+4. totals summarizes task count by modality, material count, shot count, and total duration. Never calculate, guess, or display a price.
+5. approved must be false when unresolved is non-empty, reference use is unconfirmed, the production plan is unconfirmed, or a required generation parameter is missing.
+6. Initial status is exactly blocked or ready; never mark a shot ready with incomplete dependencies.
+7. prompt_profile is exactly seedance2 or h3. Final generation uses character images, a temporal storyboard, a shot board, an audio reference, and the video prompt. Add Blender previz only when the two still boards cannot prove a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move. Seedance 2.0 uses `full-reference`, and H3 uses `Ref2VA`. If the Provider cannot carry the inputs this shot actually requires, mark unresolved; never silently omit one.
+8. When the Provider supports MiniMax H3 native audio (video.native-audio), default audio_strategy.mode to native and record a timed speech_timeline, one cross-shot music_anchor, and paired adjacent transition_pair values; do not schedule separate TTS for those shots. Use post-dub/independent only when the user explicitly requests dubbing or the Provider/model is confirmed to lack native audio; do not silently fall back to dubbing just because capability is unknown. Adjacent H3 transition_out and transition_in values must match as a pair.
+9. Every shot uses image_strategy.mode=generate, board_type=shot-board, and panel_grid_size=2..16 (default 6, adjusted to actual beats), producing independent temporal-storyboard and shot-board assets. Standard projects default to storyboard_strategy.mode=image and previz_strategy={"mode":"none"}. Use storyboard_strategy.mode=blender with previz_strategy={"mode":"blender","purpose":"motion-reference","fps":12} only when the two boards still cannot unambiguously express a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move; name that spatial risk in review_checks. Only StarRouter MiniMax-H3 currently accepts the Blender motion reference through TempFile.org hosting; otherwise mark unresolved. Do not bind future asset versions. Composition elements require source, selected-asset, or confirmed-blocking evidence; block unsupported elements in unresolved. overflow_strategy is exactly compose-assets or reject.
+9a. Before generation, realize `required_board_types`: a video shot must generate temporal and shot-board; if the preflight primary type is narrative, blocking, action, choreography, comprehensive, or director-track, also generate that primary type; if it is previz-3d, it is an additional spatial reference and never replaces temporal or shot-board. `image_strategy.board_type=shot-board` is only a compatibility field for the executor and must not override `primary_board_type`.
+10. Tailor review_checks to actual shot risk; every item must be a concrete, named, observable phrase (e.g. "mirror reflection consistency", "contact point readable through rain") derived from the contract's ai_risks and evidence elements: dialogue checks speaker identity and every visible mouth state; group shots check headcount, order, facing, and limb ownership; prop interaction checks before/contact/after states and adjacent objects; walking or fights check foot contact, weight, direction, speed, and pose across cuts; in-frame screens check whether precise post-production graphics are required. When composition risks apply, turn each into an explicit check: mirrors/glossy water verify the reflection matches the real person and actual dressing, with no extra fingers, camera, or reflection anomaly; rain, fog, neon, and smoke/dust verify the weather and colored light are evidenced, directions are consistent, and neither buries the contact point nor identity; crowd shots verify headcount, identity, costume-color anchors, and no extra bystanders; weapons and multi-hand interaction verify prop form, holding relationship, hand count, and before/contact/after states; a narrow door gap or foreground occlusion verifies visible width, both eyes intact, and the key action uncovered; high-speed trails verify both eyes, the face, and the contact point stay sharp and the subject is not a blob; front-and-back pincer/encirclement verifies both threats' identity, distance, and facing stay readable in foreground and background; vertical shots verify faces, key props, and contact points sit in the upper-middle safe area, the bottom subtitle safe area is clear, and motion edges keep a 10% margin. Keep only the risks this shot actually has. Never copy one generic checklist onto every shot.
+
+11. Only a shot whose previz_strategy.mode is blender runs direct-blender-previz before generate-blender-previz renders and scores it. Previz never replaces the temporal storyboard, shot board, character, scene, prop, wardrobe, material, or final-look references.
+
+
+13. `video_strategy.visible_character_keys` lists every `char-*` that is actually visible in the shot and exactly matches the asset keys in storyboard `characters[].appearance`; an off-screen speaker is included only when visible. Every listed key binds one selected image in reference_assets with `role=character_identity`; never bind only the protagonist or speaker. Use an empty array for a shot with no visible character. `video_strategy.visible_event_keys` lists every stable visible event key that may produce action sound, or an empty array when none exists. Final video-prompts bind one temporal storyboard, one shot board, and one audio reference. Bind exactly one Blender reference only when the plan enables it. An asset board is optional and never replaces either a character image or either storyboard.
+
+14. Every native-audio ambience item carries `evidence:{source,detail}`, where source is script, storyboard, scene_asset, or audio_reference. Every action_sounds item carries an `evidence_event_key` that exactly references this shot's `visible_event_keys`. Never prescribe footsteps without a visible walking event, energy sound without a visible energy event, or generic "cinematic sound" without evidence.
+
+No API key, token, remote business ID, Markdown, or explanation.

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

@@ -0,0 +1,52 @@
+# Codex 执行合同：production_plan
+
+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：短剧制片主任。把导演本、已选资产、已选分镜、Provider 能力和用户预算限制汇总为逐镜执行计划；不得提交任务。
+
+【导演本】
+{director_book}
+
+【已选分镜】
+{storyboard}
+
+【本地资产账本】
+{asset_ledger}
+
+【Provider 能力目录】
+{provider_catalog}
+
+【用户限制与已确认参数】
+{user_limits}
+
+只返回严格 JSON：{"episode_key":"...","source_versions":{},"shots":[...],"totals":{},"unresolved":[],"approved":false}。
+
+
+前置选择门：在决定 image_strategy、storyboard_strategy、Provider 或付费任务之前，先逐镜阅读导演本、已选分镜和用户限制，判断本镜主要不确定性属于：单一构图/首尾状态（single-frame）、连续动作与精确状态（temporal）、剧情因果/发现/情绪转折（narrative）、人物路线与空间调度（blocking）、格斗/追逐/冲击动作链（action）、舞蹈/武术/运动姿态拆解（choreography）、场景鸟瞰（scene-overview）、平面规划（scene-plan）、同一场景多视角固定（scene-turnaround）、导演节奏交付（director-track）或高风险全维度主镜头（comprehensive）。只选择能解决当前主要风险的最小主类型，并在 `selection_basis` 写出命中的原文证据和排除其他类型的理由；不得因为默认值或模型习惯直接把所有镜头标成 shot-board。
+
+完整板型合同：standard 视频镜头的 `required_board_types` 至少包含 `temporal` 与 `shot-board`；主类型为 `narrative`、`blocking`、`action`、`choreography`、`comprehensive` 或 `director-track` 时，保留双板并加入主类型；空间风险无法由静态双板证明时追加 `previz-3d`，不得用它替代 temporal 或 shot-board。single-frame、scene-overview、scene-plan、scene-turnaround 等非视频资产不自动追加双板。`image_strategy.board_type` 继续固定为 `shot-board` 以兼容现有图片生成执行器，但它不代表前置选择的主故事版类型。
+
+规则：
+1. 只能引用账本中存在的本地稳定 key 和 selected version；缺失或错版写入 unresolved。
+2. 只选择 Provider 声明支持的模态和参数；不得猜测模型或工作流，不计算或展示具体价格。
+3. 文本提示词由 Codex 生成，不列入外部文本模型调用。
+4. totals 只汇总各模态任务数、素材数量、镜头数和总时长，不包含价格或费用估算。
+5. unresolved 非空、参考素材用途未确认或任何制作参数缺失时，approved 必须为 false。
+6. status 初始只能是 blocked 或 ready；依赖不齐不得写 ready。
+8. Provider 支持 MiniMax H3 原生音频（video.native-audio）时，audio_strategy.mode 默认写 native，并记录逐秒 speech_timeline、跨镜统一 music_anchor 和相邻镜 transition_pair；这类镜头不得另排 TTS。仅当用户明确要求后配、或 Provider/模型明确声明不支持原生音频时才写 post-dub/independent 并排入独立音频流程；能力未明确时默认按原生处理，不得因信息缺失而静默转后配。相邻 H3 镜头的 transition_out 与 transition_in 必须成对一致。
+9a. 生成前必须先落实 `required_board_types`：视频镜头至少生成 temporal 与 shot-board；若前置选择的是 narrative、blocking、action、choreography、comprehensive 或 director-track，则额外生成对应主类型；若选择 previz-3d，则它只是附加空间参考，不能替代 temporal 或 shot-board。`image_strategy.board_type=shot-board` 只是执行器兼容字段，不得覆盖 `primary_board_type`。
+10. review_checks 按逐镜真实风险填写，每个检查项必须是具体、具名、可观察的短语（如「镜面反射一致性」「雨幕接触点可读」），并从合同 ai_risks 与证据元素反推：对白检查说话人与每张可见嘴部状态；多人镜检查人数、顺序、朝向和肢体归属；道具交互检查接触前/中/后状态及邻接物；步行/打斗检查落脚、重心、方向、速度和跨镜姿态；画内屏幕检查是否需要后期精确图形。命中构图风险时必须逐项落为检查：镜面/光滑水面核对映像与真人及现场陈设一致、无多指、无多余摄像机或反射穿帮；雨、雾、霓虹、烟尘核对天气光色有据、雨丝/烟向统一且不淹没接触点与人物身份；群体镜核对人数、身份、服色锚点且无多余路人；武器与多手交互核对形制、持有关系、手数与接触前/中/后；窄门缝或前景遮挡核对可见宽度、双眼完整与关键动作不被盖；高速拖影核对双眼、面部与接触点清晰可辨、主体不糊；前后夹击/包围核对前后景两方身份、距离与朝向可辨；竖屏镜核对脸、关键道具与接触点在中上安全区、底部字幕安全区不被占、动势边缘留 10% 余量。只保留本镜命中的风险项，不得复制同一组空泛检查覆盖所有镜头。
+
+11. 仅当 previz_strategy.mode=blender 时，白模才由 direct-blender-previz 先生成导演合同，再由 generate-blender-previz 生成和评分；白模不能替代时间故事版、镜头分镜板、人物、场景、道具、服装、材质或最终画质参考。
+
+
+
+14. 原生声音计划中的每条 ambience 必须携带 `evidence:{source,detail}`，source 只能是 script、storyboard、scene_asset 或 audio_reference；每条 action_sounds 必须携带 `evidence_event_key` 并精确引用本镜 `visible_event_keys`。禁止给没有走动事件的镜头安排脚步、给没有可见能量事件的镜头安排能量声，或用“环境音/电影感音效”等泛化描述绕过证据。
+
+禁止 API Key、Token、远程业务 ID、Markdown 和解释。
+
+复核规则：时间故事版负责连续动作、运镜与精确状态，镜头分镜板负责景别、调度、轴线与叙事节拍；两者必须是独立资产。只有双板和其他独立参考实际超过 Provider 图片槽位才使用 compose-assets。项目 automation_mode=true 且仅缺 Provider 已支持的分辨率时，直接选支持值并写入计划。
+
+standard 默认写 `storyboard_strategy.mode=image` 与 `previz_strategy={"mode":"none"}`；只有双分镜板仍无法无歧义表达复杂路线、多人交互、精确接触、轴线风险或连续运镜时，才追加可选白模作为 Blender 运动参考。
+
+每镜必须明确列出 `visible_character_keys`、`visible_event_keys`，并让声音事件携带 `evidence_event_key`。

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

@@ -0,0 +1,37 @@
+# Codex execution contract: continuity_plan
+
+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 current selected project documents and reviewed assets.
+
+Execution role: script supervisor and continuity director. Build shot-level spatial and state handoffs from the approved structured storyboard, director book, production plan, and selected assets. Do not generate media or refer to dual boards or previz that do not exist yet.
+
+[EPISODE]
+{episode_key}
+
+[SELECTED STORYBOARD]
+{storyboard_json}
+
+[SELECTED DIRECTOR BOOK]
+{director_book_json}
+
+[SELECTED PRODUCTION PLAN]
+{production_plan_json}
+
+[SELECTED ASSETS]
+{selected_assets_json}
+
+Return strict JSON only. The top-level fields are exactly `episode_key,source_versions,scenes,shots,unresolved,approved`. `source_versions` contains exactly the three vNNN versions `storyboard,director-book,production-plan`.
+
+Each scene contains exactly `scene_key,coordinate_mode,axis_id,axis_description,camera_side,anchors,lighting_anchor`. Each shot contains exactly `shot_number,scene_key,camera_setup_id,start_state,end_state,transition_link,inherited_fields,allowed_changes,evidence`.
+
+start_state and end_state contain exactly `actors,props,axis_id,camera_side,lighting_anchor`. An actor contains exactly `actor_key,zone,depth,facing,eyeline_target,screen_direction,entry_edge,exit_edge,posture,held_props`; a prop contains exactly `prop_key,zone,depth,state,held_by`. transition_link contains exactly `mode,source_shot_number,source_camera_setup_id,enabled,reason,required_provider_capability`.
+
+Rules:
+1. `coordinate_mode` is always `semantic`; zone and depth use screen-space and depth-layer semantics. Never estimate exact metric coordinates. If later dual boards or previz expose a conflict, return upstream and create a new continuity-plan version.
+2. The first shot is independent with null source/capability fields and enabled=false. Compare the previous end_state with the current start_state, then call `recommendTailLink`; use previous-tail only for the same scene and camera_setup_id, the same axis/camera side, and no unexplained difference.
+3. previous-tail has enabled=true, source_shot_number equal to the immediately preceding shot, source_camera_setup_id equal to that shot's setup, and required_provider_capability exactly `video.first-frame`. A camera change, axis crossing, scene change, or intentional discontinuity is independent.
+4. inherited_fields lists only fields actually inherited. allowed_changes accepts only changes supported by the script, director book, or approved blocking; never use it to hide a conflict.
+5. Never infer an undeclared prop, held-prop relationship, entrance, exit, absent character, or extra character. Put every unsupported state, coordinate, or change in unresolved; each item is exactly `{id,question,affects}`.
+6. evidence contains at least one verifiable source. approved=false whenever unresolved is non-empty; approved=true only when every shot is continuous and every source version is currently selected.
+7. Read and preserve storyboard shot_group intent: a group id is contiguous and keeps one type/pattern, but it is narrative organization rather than an automatic tail chain. Montage, reverse angle, camera change, axis crossing, or scene change remains independent. Only a consecutive previous-tail chain becomes a Runtime generation group, executed serially as reviewed selection → tail extraction → next-shot first frame.
+
+No Markdown, explanation, API key, token, temporary URL, or path outside the project.

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

@@ -0,0 +1,37 @@
+# Codex 执行合同：continuity_plan
+
+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。花括号字段是当前项目 selected 文档和已审核资产组装的具名输入槽。
+
+执行职责：场记与连续性导演。依据已批准结构化分镜、导演本、制作计划和选中资产，建立逐镜空间与状态承接；不生成媒体，也不引用尚未生成的双分镜板或白模。
+
+【分集】
+{episode_key}
+
+【已选分镜】
+{storyboard_json}
+
+【已选导演本】
+{director_book_json}
+
+【已选制作计划】
+{production_plan_json}
+
+【已选资产】
+{selected_assets_json}
+
+只返回严格 JSON，顶层字段必须且只能是 `episode_key,source_versions,scenes,shots,unresolved,approved`。`source_versions` 必须且只能包含 `storyboard,director-book,production-plan` 三个 vNNN 版本。
+
+每个 scene 必须且只能包含：`scene_key,coordinate_mode,axis_id,axis_description,camera_side,anchors,lighting_anchor`。每个 shot 必须且只能包含：`shot_number,scene_key,camera_setup_id,start_state,end_state,transition_link,inherited_fields,allowed_changes,evidence`。
+
+start_state 与 end_state 必须且只能包含 `actors,props,axis_id,camera_side,lighting_anchor`。actor 必须且只能包含 `actor_key,zone,depth,facing,eyeline_target,screen_direction,entry_edge,exit_edge,posture,held_props`；prop 必须且只能包含 `prop_key,zone,depth,state,held_by`。transition_link 必须且只能包含 `mode,source_shot_number,source_camera_setup_id,enabled,reason,required_provider_capability`。
+
+规则：
+1. `coordinate_mode` 固定为 `semantic`，zone/depth 使用画面和景深语义位置；不得估算精确米制坐标。后续双分镜板或白模暴露冲突时，必须回退并生成连续性计划新版本。
+2. 第一镜必须为 independent，来源字段和能力为 null，enabled=false。相邻镜先比较上一镜 end_state 与本镜 start_state，再调用 `recommendTailLink`；只有同场景、同 camera_setup_id、同轴线/机位侧且无未解释差异时才用 previous-tail。
+3. previous-tail 必须 enabled=true，source_shot_number 是紧邻上一镜，source_camera_setup_id 与上一镜一致，required_provider_capability 固定为 `video.first-frame`。切机位、越轴、换场或有意跳变必须 independent。
+4. inherited_fields 只列实际继承字段。allowed_changes 只接受剧本、导演本或已批准调度明确支持的变化；不得用它掩盖冲突。
+5. 不得推断未声明的道具、持物、入口、出口、人物缺席或新增人物。缺证据的状态、坐标和变化逐项写入 unresolved；每项严格为 `{id,question,affects}`。
+6. evidence 至少包含一个可核查来源。unresolved 非空时 approved=false；只有全部镜头连续且来源版本都是当前 selected 时 approved=true。
+7. 读取并尊重分镜中的 shot_group：同组 id 必须连续且 type/pattern 不变，但它是叙事组织，不自动等于尾帧链。蒙太奇、反打、切机位、越轴或换场仍必须 independent；只有连续 previous-tail 镜头才构成 Runtime 的生成镜头组，并按前镜审核选版→提取尾帧→下一镜首帧的顺序串行执行。
+
+禁止 Markdown、解释、API Key、Token、临时 URL 和项目外路径。

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/review-drama-script/prompts/review_drama_script.en.txt

@@ -0,0 +1,28 @@
+# Codex execution contract: review_drama_script
+
+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 script review editor. Review only; do not rewrite the script or call an external text model.
+
+[SCRIPT]
+{script}
+
+[SOURCE ANALYSIS]
+{source_analysis}
+
+[CREATIVE BRIEF]
+{brief}
+
+[DRAMA BIBLE]
+{bible}
+
+[EPISODE OUTLINE]
+{outline}
+
+[COMPLIANCE RULES]
+{compliance_rules}
+
+Check opening pull, scene rhythm, conflict and payoff, character-specific dialogue, factual and version continuity, compliance red lines, target duration, and production feasibility. Treat duration as a planning target rather than an exact cutoff: natural variance of about ±15% is not a defect; beyond that, require a narrative reason and user confirmation instead of mechanical trimming. Production feasibility also checks that each action beat is singular, visible, and reliably performable by a video model; that extended dialogue has necessary action, reaction, or intentional pauses; and that position, presence, and held props remain continuous across neighboring beats. A faithful conversion must not be asked to invent unsupported action merely to satisfy these checks. Each issue must include scene_or_line, severity (P0/P1/P2), dimension, finding, and required_change. Do not substitute vague ratings for evidence. approved must be false if any P0/P1 issue exists or compliance fails.
+
+Return strict JSON only, with no Markdown or extra keys:
+{"episode_key":"ep-001","script_version":"v001","dimensions":{"opening":{"status":"passed","note":""},"pace":{"status":"passed","note":""},"payoff":{"status":"passed","note":""},"dialogue":{"status":"passed","note":""},"continuity":{"status":"passed","note":""},"production_feasibility":{"status":"passed","note":""}},"issues":[],"compliance":{"passed":true,"issues":[]},"approved":true}

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/review-drama-script/prompts/review_drama_script.zh.txt

@@ -0,0 +1,28 @@
+# Codex 执行合同：review_drama_script
+
+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：短剧剧本复核编辑。只审查，不改写剧本；不得调用外部文本模型。
+
+【待审剧本】
+{script}
+
+【来源分析】
+{source_analysis}
+
+【创作简报】
+{brief}
+
+【故事圣经】
+{bible}
+
+【分集目录】
+{outline}
+
+【合规清单】
+{compliance_rules}
+
+逐项核对开场抓力、场内节奏、冲突与兑现、角色化对白、事实与版本连续性、合规红线、目标时长和生成可执行性。时长按策划目标而非精确卡秒评估；约 ±15% 内的自然偏差不作为缺陷，超过时检查是否有动作、对白或钩子的叙事理由并要求用户确认，而不是机械删减。生成可执行性还要检查动作节拍是否单一、可见、可由视频模型稳定表现，连续对白是否有必要的动作/反应/停顿，以及人物位置、在场关系和道具持有是否在相邻节拍间连续；忠实转换不得因这些检查被要求补造来源没有的动作。问题必须给出 scene_or_line、severity（P0/P1/P2）、dimension、finding、required_change；不得用笼统评价代替证据。P0/P1 或合规失败时 approved 必须为 false。
+
+只返回严格 JSON，不要 Markdown 或额外键：
+{"episode_key":"ep-001","script_version":"v001","dimensions":{"opening":{"status":"passed","note":""},"pace":{"status":"passed","note":""},"payoff":{"status":"passed","note":""},"dialogue":{"status":"passed","note":""},"continuity":{"status":"passed","note":""},"production_feasibility":{"status":"passed","note":""}},"issues":[],"compliance":{"passed":true,"issues":[]},"approved":true}

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/revise-drama-storyboards/prompts/agent_shot_variant_analysis.en.txt

@@ -0,0 +1,57 @@
+# Codex execution contract: agent_shot_variant_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: shot variant analysis expert.
+Analyze the current shot and provide multiple strong variant ideas.
+
+{cinematography_vocabulary}
+
+Current shot description:
+{panel_description}
+
+Current shot_type:
+{shot_type}
+
+Current camera_move:
+{camera_move}
+
+Location:
+{location}
+
+Characters:
+{characters_info}
+
+Task:
+Generate at least 3 shot variants while preserving narrative continuity, character identity, and location consistency.
+
+Composition-defect repair mode:
+When the input includes a failed review, a composition defect, or a composition contract (visual_plan.composition, composition_contract, photographyPlan.composition), the variants must repair the defect under these rules:
+1. Change exactly one variable per variant and begin description with "Changed variable: X; Repairs: Y". The fixed priority of variables is: (1) subject placement (left/right, foreground/midground/background, camera height); (2) eyeline direction and negative space; (3) foreground/midground/background layers and occlusion; (4) light direction and color temperature; (5) action split (break an unreadable action into anticipation-contact-result-reaction phases, or add a reaction shot). Shot scale, angle, and camera move are the last resort.
+2. The shot's single primary readable fact, its primary/secondary composition codes, and the axis and screen direction shared with neighboring shots stay unchanged by default; a direction fix is allowed only when the defect itself is an axis/screen-direction break and the user explicitly requests it, with the basis stated in description.
+3. Delete unsupported elements first: rain, snow, fog, neon, mirror/glossy-water reflection, smoke/dust, doors/windows/columns, weapons, or extra figures that appear in the image but lack support in the source text, selected scene/prop assets, or confirmed blocking must be removed or replaced with composition-only geometry achievable through camera position, blocking, and exposure; never add or strengthen them to make the shot "more cinematic" or "more atmospheric". A variant needing set_dressing, weather_atmosphere, or action_blocking evidence (each component of a + compound) must cite the support in description; if you cannot cite it, do not propose the variant.
+4. Every repair variant must be verifiable: description names the contract field it repairs (read_order, subject_placement, eyeline_room, depth_layers, light_and_color_geometry, action_vector, tail_state) or the risk code (see the risk table in section 0.4 of references/director-composition.md), and states what the viewer should read at first glance after the repair.
+5. Creative exploration without a defect may still vary angle, scale, timing, and atmosphere, but it likewise adds no unsupported elements, reselects no codes, and breaks no continuity with neighboring shots.
+
+Output format (JSON array only):
+[
+  {
+    "id": 1,
+    "title": "Variant title",
+    "description": "What changes and why it works",
+    "shot_type": "target shot type",
+    "camera_move": "target camera move",
+    "video_prompt": "short motion-focused prompt",
+    "creative_score": 8.5
+  }
+]
+
+Rules:
+1. Provide at least 3 variants.
+2. Keep each variant practically producible.
+3. creative_score range: 0-10.
+4. A defect-repair variant changes exactly one composition variable and states "Changed variable: X; Repairs: Y" in description; never change several variables in one variant.
+5. Never add or strengthen rain, snow, fog, neon, mirrors, smoke, doors/windows/columns, weapons, or extra figures to make the shot "more cinematic"; unsupported elements of those kinds must be deleted, not embellished, and a variant requiring set_dressing/weather_atmosphere/action_blocking evidence (including a + compound) without citable support must not be proposed.
+6. Do not reselect the primary/secondary composition codes or flip the axis or neighboring screen direction unless the defect itself is that direction error and the user explicitly requests it.
+7. Keep JSON strict and valid.
+8. ⚠️ JSON SAFETY: All quotation marks in text MUST be converted to corner brackets「」in JSON string values. NEVER use raw ASCII double quotes " inside string values.

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/revise-drama-storyboards/prompts/agent_shot_variant_analysis.zh.txt

@@ -0,0 +1,169 @@
+# Codex 执行合同：agent_shot_variant_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：专业的电影分镜分析师。你的任务是分析一个镜头画面，并推荐多种有创意的镜头变体方案。
+
+{cinematography_vocabulary}
+
+======================================
+【输入信息】
+======================================
+
+## 当前镜头描述
+{panel_description}
+
+## 景别与运镜
+景别: {shot_type}
+运镜: {camera_move}
+场景: {location}
+
+## 出场角色
+{characters_info}
+
+## 当前画面
+（附带参考图片）
+
+======================================
+【分析任务】
+======================================
+
+请分析当前镜头画面内容，推荐 **5-8 个** 多样化的镜头变体方案。
+
+变体类型应覆盖以下维度（不必全部使用，选择最适合当前画面的）：
+
+1. **视角变换**
+   - 正反打：如果画面是 A 看 B，可以改为 B 看 A
+   - 主观视角：某角色的第一人称视角（如看手机屏幕、看窗外）
+   - 俯拍/仰拍：改变拍摄角度
+
+2. **景别变化**
+   - 拉远：从特写扩展到中景/全景，展示环境关系
+   - 推近：从中景聚焦到特写，强调情绪/细节
+   - 局部特写：聚焦画面中的某个物品或身体部位
+
+3. **时间/动作变化**
+   - 动作前/后：展示动作发生前一刻或后一刻
+   - 反应镜头：另一角色对当前画面的反应
+
+4. **场景/氛围变化**
+   - 环境镜头：聚焦背景氛围（窗外、室内布置）
+   - 光影变化：不同光线氛围（如逆光、剪影）
+
+======================================
+【构图缺陷修复模式】
+======================================
+
+当输入附带审片失败、构图缺陷描述，或分镜含 visual_plan.composition、composition_contract、photographyPlan.composition 合同时，推荐方案以修复缺陷为目的，必须遵守：
+
+1. 一次只改变一个变量，并在 description 开头写明「改变变量：X；修复：Y」。变量调整优先级固定为：①主体位置（水平左右、前中后景、机位高度）②视线方向与留白 ③前中后景层次与遮挡 ④光源方向与色温 ⑤动作拆分（把不可读动作拆成蓄力—接触—结果—反应阶段，或补一个反应镜头）；景别、拍摄角度与运镜是最后才动的变量。
+2. 本镜唯一首要可读信息、primary/secondary composition code、与相邻镜的轴线和屏幕方向默认保持不变；只有当缺陷本身就是越轴或方向翻转、且用户明确要求时才允许修正方向，并在 description 说明依据。
+3. 无证据元素删除优先：雨、雪、雾、霓虹、镜面/水面反射、烟尘、门窗廊柱、武器、多余人物若在原画面出现却在原文、已选场景/道具资产或已确认调度中找不到依据，修复方向是删除或替换为只靠机位、调度、曝光成立的 composition-only 几何；禁止以「更电影感」「更有氛围」为由新增或强化这些元素。需要 set_dressing、weather_atmosphere、action_blocking（含以 + 连接的复合证据）支撑的变体，必须在 description 注明依据来源，写不出依据就不要推荐该变体。
+4. 每个修复变体必须可验证：description 写明它修复的是哪个合同字段（read_order、subject_placement、eyeline_room、depth_layers、light_and_color_geometry、action_vector、tail_state）或哪个风险码（见 references/director-composition.md 0.4 风险码表），以及修复后第一眼应读到什么。
+5. 无缺陷的创意探索仍可覆盖上节四类维度，但同样不得新增无证据元素、不得改选 code、不得翻转与相邻镜的连续性。
+
+======================================
+【输出格式】
+======================================
+
+返回 JSON 数组，每个推荐包含：
+
+```json
+[
+  {
+    "id": 1,
+    "title": "简短标题（如：主观视角-手机屏幕）",
+    "description": "详细描述这个镜头变体会呈现什么画面",
+    "shot_type": "推荐景别（如：主观特写）",
+    "camera_move": "推荐运镜（如：固定）",
+    "video_prompt": "用于图片生成的详细提示词，使用年龄+性别描述人物",
+    "creative_score": 5
+  }
+]
+```
+
+**字段说明**：
+- `id`: 序号（1-8）
+- `title`: 简短标题，格式如"类型-具体内容"
+- `description`: 详细描述该变体的画面内容
+- `shot_type`: 推荐景别，如"平视中景"、"俯拍全景"、"主观特写"
+- `camera_move`: 推荐运镜，如"固定"、"缓推"
+- `video_prompt`: 图片生成提示词，必须用"年龄+性别"替代角色名（如"年轻女子"、"中年男子"）
+- `creative_score`: 创意程度 1-5，5为最有创意
+
+======================================
+【示例】
+======================================
+
+**输入画面**: 年轻女子躺在床上看手机
+
+**推荐变体**:
+```json
+[
+  {
+    "id": 1,
+    "title": "主观视角-手机屏幕",
+    "description": "第一人称视角，展示手机屏幕内容，手机边缘和手指可见",
+    "shot_type": "主观特写",
+    "camera_move": "固定",
+    "video_prompt": "POV shot of a smartphone screen, fingers holding the phone edges, bright screen glow in dark room, close-up perspective",
+    "creative_score": 5
+  },
+  {
+    "id": 2,
+    "title": "脸部特写",
+    "description": "女子脸部特写，手机屏光打在脸上，呈现表情细节",
+    "shot_type": "特写",
+    "camera_move": "固定",
+    "video_prompt": "Close-up of a young woman's face illuminated by phone screen light, soft blue glow on skin, expression of focus, lying down angle",
+    "creative_score": 4
+  },
+  {
+    "id": 3,
+    "title": "俯拍全景",
+    "description": "从天花板角度俯拍，展示整个床铺和女子姿态",
+    "shot_type": "俯拍全景",
+    "camera_move": "固定",
+    "video_prompt": "Top-down bird's eye view of bedroom, young woman lying on bed holding phone, cozy blankets, bedroom interior visible",
+    "creative_score": 4
+  },
+  {
+    "id": 4,
+    "title": "手部特写",
+    "description": "特写手指在手机屏幕上滑动",
+    "shot_type": "极端特写",
+    "camera_move": "固定",
+    "video_prompt": "Extreme close-up of feminine fingers swiping on smartphone touchscreen, colorful app interface, shallow depth of field",
+    "creative_score": 3
+  },
+  {
+    "id": 5,
+    "title": "侧脸剪影",
+    "description": "逆光侧拍，女子轮廓剪影，手机屏幕微微发光",
+    "shot_type": "近景",
+    "camera_move": "固定",
+    "video_prompt": "Silhouette profile of young woman in dark room, backlit by dim blue phone glow, artistic dramatic lighting, moody atmosphere",
+    "creative_score": 5
+  }
+]
+```
+
+======================================
+【禁止规则】
+======================================
+
+❌ 推荐与当前镜头完全相同的方案
+❌ 修复缺陷的一个变体同时改变多个构图变量；不写「改变变量：X；修复：Y」
+❌ 使用角色名，必须用"年龄段+性别"（年轻女子、中年男子等）
+❌ 以「更电影感/更有氛围」为由新增或强化雨、雪、雾、霓虹、镜面、烟尘、门窗廊柱、武器、多余人物；无证据的这类元素必须删除而不是美化，需要 set_dressing/weather_atmosphere/action_blocking（含 + 复合）证据却写不出依据的变体不得推荐
+❌ 改选 primary/secondary composition code、翻转轴线或相邻镜屏幕方向（缺陷本身就是方向错误且用户明确要求除外）
+❌ 推荐超出场景合理性的内容（如室内镜头推荐户外场景）
+❌ 推荐少于 5 个或多于 8 个变体
+❌ video_prompt 使用中文（必须英文）
+
+======================================
+【输出】
+======================================
+
+只返回 JSON 数组，不需要 markdown 代码块。
+⚠️ JSON安全：所有引号（""''等）在 JSON 字符串值中必须统一替换为「」，严禁出现未转义的英文双引号 "。

package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/revise-drama-storyboards/prompts/agent_shot_variant_generate.en.txt

@@ -0,0 +1,64 @@
+# Codex execution contract: agent_shot_variant_generate
+
+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: storyboard image generation assistant.
+Generate one new variant image that keeps identity/style continuity while applying the requested camera variation.
+
+Reference context:
+- Original description: {original_description}
+- Original shot type: {original_shot_type}
+- Original camera move: {original_camera_move}
+- Location: {location}
+- Characters: {characters_info}
+
+Variant request:
+- Variant title: {variant_title}
+- Variant description: {variant_description}
+- Target shot type: {target_shot_type}
+- Target camera move: {target_camera_move}
+
+[COMPOSITION-DEFECT REPAIR DISCIPLINE — MANDATORY]
+1. When the variant is a composition-defect repair, change exactly the one variable named by the variant request (one of subject placement; eyeline and negative space; foreground/midground/background layers and occlusion; light direction and color temperature; action split). Every other item — subject placement, eyeline direction, layer contents, key-light direction, screen direction and axis, tail pose — must stay consistent with the reference images and the composition contract (visual_plan.composition / composition_contract / photographyPlan.composition). A creative (non-repair) variant follows {target_shot_type}/{target_camera_move} and the variant request to vary angle, shot scale, composition, or atmosphere, but rules 2–4 still bind it: the evidence gate, deletion over embellishment, and the headcount/identity/weapon/occlusion/vertical-safe-area preservation rules all apply.
+2. Deletion beats embellishment: rain, snow, fog, neon, mirror/glossy-water reflection, smoke/dust, doors/windows/columns, weapons, or extra figures that have no basis in the source text, selected scene/prop assets, or confirmed blocking must be removed or replaced with composition-only imagery achievable through camera position, blocking, and exposure; never strengthen them to make the variant "more cinematic" or "more atmospheric".
+3. An element requiring set_dressing, weather_atmosphere, or action_blocking evidence (every component of a + compound) may be drawn only when supported by the original shot description, character/location references, or the confirmed variant request; with no support anywhere, do not draw it.
+4. Unless the variant asks otherwise, keep crowd headcount, positions, facing, eyeline network, and costume-color anchors unchanged; weapon form, holding relationship, and contact point must not drift; occlusion never covers the face or key action; in 9:16 keep both eyes, faces, key props, and contact points in the upper-middle safe area, keep the bottom subtitle safe area clear, and leave at least a 10% margin at motion edges.
+
+Generation prompt seed:
+{video_prompt}
+
+Character assets:
+{character_assets}
+
+Location asset:
+{location_asset}
+
+Output aspect ratio:
+{aspect_ratio}
+
+[⚠️⚠️⚠️ GLOBAL ART-STYLE CONSISTENCY — HIGHEST PRIORITY — NON-NEGOTIABLE]
+Every element in the frame MUST be rendered in ONE unified art style. Style mixing is FORBIDDEN:
+- Main characters, supporting characters, passersby, crowd, and ANY background figures MUST all share the EXACT same art style, line work, coloring method, and lighting logic as the main character references.
+- Passersby / NPCs without their own reference MUST still be painted in the same style as the main characters — same outline weight, same shading, same rendering texture. NEVER improvise a different style for them.
+- Environments, props, sky, vegetation, and architecture MUST match the rendering dimension and texture of the character/location references.
+- FORBIDDEN: 2D anime characters on 3D photoreal backgrounds; hand-drawn people next to photo-textured props; any visual mismatch.
+- If references are clearly 2D (hand-drawn / anime / comic / illustration), the ENTIRE frame MUST be 2D — absolutely NO 3D, CG, CGI, Blender, Pixar/Disney 3D-cartoon look, Unreal-Engine render, photorealism, plastic skin, or 3D-model feel.
+- If references are clearly photoreal / live-action, keep the whole frame consistently photoreal; do not anime-ify or cartoon-ify anyone.
+- Line thickness, cel-vs-gradient shading, shadow hardness, and highlight treatment MUST be uniform.
+- Anti-AI look: avoid over-sharpening, overdone HDR, oversaturation, plastic/wax skin, greasy textures, extra fingers, distorted faces; keep natural linework/brushwork and coherent lighting across the whole frame.
+
+Style requirement (treat as authoritative):
+{style}
+
+[PROJECT VISUAL BIBLE — independent from art style, must be followed]
+{visual_bible}
+
+Execution rules:
+1. Preserve character identity and outfit continuity unless the variant asks otherwise.
+2. Preserve location continuity.
+3. Change framing/angle/composition according to target shot and camera move.
+4. If characters_info or location_asset includes fixed slots / available slots, keep every visible character anchored to the same fixed slot instead of drifting to another area.
+5. Keep one-frame output only, no text, watermarks, numbering, dialogue bubbles, or UI elements.
+6. Passersby / crowd / supporting characters without dedicated references MUST still be drawn in the SAME style as the main character references.
+7. No 2D/3D mixing; no multiple rendering textures within one frame.
+8. Follow the composition-defect repair discipline: a repair variant changes only the single variable the variant names; a creative variant may vary angle/scale/composition per {target_shot_type}/{target_camera_move}, but every variant must delete unsupported weather, neon, mirrors, smoke, doors/windows/columns, weapons, or extra figures or replace them with composition-only geometry, never adding or strengthening them for cinematic effect.