Skip to content
Back to search

manju-director

SkillWorks with Claude CodeGitHub

"Create and revise AI comic-drama or short-drama projects through a persistent, evidence-gated ReAct loop: scripts, assets, shots, model inputs and observed reviews. Use for production, an existing project's next shot, or maintenance of this workflow."

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/liangie7420--master-director/badge.svg)](/listings/liangie7420--master-director)

HTML

<a href="/listings/liangie7420--master-director"><img src="/api/listings/liangie7420--master-director/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 75 lines
  1. ---
  2. name: manju-director
  3. description: "Create and revise AI comic-drama or short-drama projects through a persistent, evidence-gated ReAct loop: scripts, assets, shots, model inputs and observed reviews. Use for production, an existing project's next shot, or maintenance of this workflow."
  4. ---
  5. # Manju Director / 漫剧导演
  6. Run the whole skill as **Reason → Act → Observe → Reflect → next action**. Story, assets, storyboards and prompts are work products inside this loop. A phase checklist, a long prompt or a written plan does not substitute for execution and observation.
  7. Use Simplified Chinese for conversation and scripts. Model-input language follows the project. Preserve original names, dialogue, approved identities and user decisions.
  8. ## Mandatory execution entry
  9. Resolve the skill root from the installed SKILL.md path. Run its `scripts/react_loop.py` by absolute path; use an explicit creative project directory. State lives in `<project>/.manju-loop/state.sqlite`, outside the installed skill.
  10. For an existing project, **resume before doing production work**:
  11. ```bash
  12. python <skill-root>/scripts/react_loop.py --project <project> resume
  13. ```
  14. For a new project, first write the scoped goal and nonempty acceptance criteria to a control JSON file, then initialize. Creating control payloads and reading the entrypoint are bootstrap work, not completed deliverables.
  15. ```bash
  16. python <skill-root>/scripts/react_loop.py --project <project> init --goal "本轮交付目标" --criteria criteria.json --mode production
  17. ```
  18. Use `revision` for an established creative revision and `maintenance` for this repository's code/document work. All modes use the same engine. Existing approval and authorization carry forward; maintenance does not open a new drama's creative approval gates.
  19. Read [the execution contract](references/react-loop.md) before the first action. The CLI's returned phase, pending action and unmet criteria determine what is legal next. If the engine cannot run, report the missing capability; do not silently fall back to a prompt-only SOP or claim the loop ran.
  20. ## Every turn
  21. 1. **Reason:** read the current state and latest user change. Select one bounded action that resolves a missing criterion or gathers necessary evidence. Record a short action reason, expected observation, target criteria and current state revision. Do not store private chain-of-thought.
  22. 2. **Act:** call `react_loop.py ... act --input action.json`. Use a registered tool; it performs the action and stores the actual result. Do not bypass the loop by calling a low-level production script directly.
  23. 3. **Observe:** read the returned logs and actual artifacts. For visual claims, use an available visual channel to inspect the exact registered media, then name the observer and coverage. Submit `observe --input observation.json` for that action ID. Summarizing the plan is not an observation. No visual capability means inconclusive/unreviewed, not passed.
  24. 4. **Reflect:** compare the observation to the criterion. Submit `reflect --input decision.json`: continue, revise, block or complete. A revision states what will change and why; the next action must record that change. Reuse supported work and reopen only affected criteria.
  25. 5. **Repeat or stop:** use the returned state. Do not start another action while observation/reflection is pending. Complete only when all scoped criteria have current evidence; blocked work retains a concrete next step.
  26. The engine enforces ordering, state revisions, real registered-tool receipts, file hashes, criterion types and completion. It cannot intercept arbitrary tools a host invokes outside this interface, verify the truth of a host's visual judgment, or make an unconfigured backend available. Do not describe those as enforced capabilities.
  27. ## Choose work from evidence, not a fixed stage order
  28. Load only the reference needed for the selected action:
  29. | Current gap | Work to perform inside the loop | Reference |
  30. |---|---|---|
  31. | Missing brief or narrative causality | Use vimax_request/story or script, process the actual request in the current host, then vimax_response; preserve user-authored material | [Script format](references/script-format.md), [bible](assets/episode-bible.md) |
  32. | Missing/unusable identity or composition | Prepare/revise character, scene, prop or keyframe material; bind real asset roles and retain prior versions | [Image engine](references/image-prompt-engine.md), relevant assets/ template |
  33. | Unclear event, geography, contact or camera | Use ViMax storyboard/decompose and take_storyboard when applicable; specify intent, visible evidence, focus, blocking, timing and start/end states | [Shot language](references/shot-language.md), [shot spec](references/shot-spec.md), [shot table](assets/shotlist-template.md) |
  34. | Invalid/ambiguous model input | Select a verified version/task and compile only necessary input; inspect actual validation result | [Video framework](references/video-prompt-framework.md), [model selection](references/model-adapters/README.md) |
  35. | Failed/uncertain continuity or performance | Use index_media/extract_frame on actual clips, inspect registered media visually, distinguish facts from semantic judgment, revise affected inputs | [Continuity](references/continuity-playbook.md), [evaluation](references/evaluation.md) |
  36. | Workflow bug or prompt experiment | Reproduce the issue, preserve baseline, change a justified factor and run checks | [Improvement work](references/agent-loop.md) |
  37. These are action choices and reference indexes, not mandatory sequential phases. Reopen creative decisions only where the current request changes them; do not repeatedly ask for already-given approval.
  38. ## Runtime tools and limits
  39. Read [integrated components](docs/INTEGRATIONS.md) when selecting ViMax, take, media, MCP or fal actions. If their optional environment is missing, install it once with `scripts/setup_integrations.py --install` as dependency bootstrap, then resume. This does not authorize paid model calls.
  40. Registered tools: `vimax_request`, `vimax_response`, `take_storyboard`, `index_media`, `extract_frame`, `fal_job_status`, `inspect`, `write_artifact`, `revise_artifact` (maintenance only), `prepare_shot`, `lint_prompt`, `run_tests`, `build_bundle` (maintenance only), `remove_artifacts` (maintenance only). See their JSON arguments in the execution contract. Existing helper scripts remain implementation tools; they do not themselves satisfy the observation/reflection requirement.
  41. Production writes use new version paths. Maintenance revisions and explicit file removals require current file hashes and retain previous bytes. Evidence changed after review becomes stale; prior acceptance cannot justify a new result. Planned states are not observed states.
  42. The MCP stdio server shares this exact core; no separate workflow is created. ViMax requests use the current host for creative responses, then the actual upstream parser validates imported responses. Preserve a pending request until a real host response exists; do not label request preparation as a completed script.
  43. This runtime has **no generation submission backend or arbitrary shell action**. The fal adapter only reads existing job status. Generation is disabled; unregistered generation actions are rejected. Deliver prepared inputs or inspect existing media, and retain an unmet visual criterion when actual video review is required. Do not mark a preparation-only goal as a completed-film goal, or drop a required subject/criterion to make completion possible.
  44. Respect existing scope and budget, including text/visual service costs. Unknown charges are not zero. On missing tools, exhausted budget, authentication problems, uncertain submission or turn limit, preserve state rather than improvising a paid workaround. `recover` records interrupted local work as unresolved and does not replay it.
  45. ## Delivery
  46. Report the goal, latest loop phase, accepted and unmet criteria, artifact paths, evidence coverage and next action. Distinguish local checks, real generation and visual acceptance. `completed` applies only to the declared goal and criterion scope, never to unobserved media quality.
  47. First-use packaging: [installation](docs/INSTALL.md). Start from the user's own story and assets.
README.md · README · 67 lines
  1. # 漫剧导演 · Manju Director
  2. 以持久化 **ReAct loop** 驱动 agent,将你的故事转为剧本、资产卡、分镜、模型输入与评审记录。
  3. **目标与状态 → 行动 → 真实结果 → 观察 → 反思 → 下一步**
  4. ## 开始使用
  5. 在源码目录安装到你使用的 AI 助手:
  6. | 助手 | 安装命令 | 使用指南 |
  7. |---|---|---|
  8. | Codex | `python scripts/manage_skill.py install --host codex` | [Codex](docs/CODEX.md) |
  9. | Claude Code | `python scripts/manage_skill.py install --host claude` | [Claude Code](docs/CLAUDE_CODE.md) |
  10. | WorkBuddy | `python scripts/manage_skill.py install --host workbuddy` | [WorkBuddy](docs/WORKBUDDY.md) |
  11. 在新会话调用 **manju-director**,提供故事、已有素材、画风、时长和目标模型版本。Agent 会先恢复或初始化项目循环,再按当前缺口推进。
  12. [安装、更新与卸载](docs/INSTALL.md) · [支持状态](docs/SUPPORT.md)
  13. ## 工作方式
  14. - **编剧与资产:**保留原故事、已确认身份和创作决定,形成可编辑的设定与资产卡。
  15. - **镜头设计:**组织可见事件、空间关系、动作节拍、运镜和起止状态。
  16. - **模型输入:**按具体任务与版本,为 Seedance、MiniMax/海螺、Google Gemini Omni、可灵准备输入。
  17. - **观察与修订:**核对实际文件与工具结果,局部修订,保留历史和证据。
  18. 创作、修订和维护共用 SQLite 状态与 JSON CLI。每次行动都必须经过观察、反思;证据变化会使验收失效,条件未满足时无法完成。[执行协议](references/react-loop.md)
  19. ## 项目结构
  20. | 入口 | 内容 |
  21. |---|---|
  22. | [SKILL.md](SKILL.md) | Agent 的执行入口 |
  23. | [references/](references/react-loop.md) | 循环协议、镜头语言、图像/视频提示词与模型适配 |
  24. | [assets/](assets/shot-spec-template.json) | 剧集、角色、场景、道具与镜头模板 |
  25. | [scripts/](scripts/react_loop.py) | 执行核心、输入检查、安装与打包工具 |
  26. | [tests/](tests/test_react_loop.py) · [benchmarks/](benchmarks/README.md) | 行为回归与六类镜头基准 |
  27. | [docs/](docs/INSTALL.md) | 平台使用、验证记录与研究接入设计 |
  28. ## 本地开发
  29. 基础循环需 Python 3.8+,仅使用标准库;扩展组件需 Python 3.10+、Node.js 与可选依赖。技能生产执行从循环入口开始;已有项目先恢复,新项目按协议声明目标与验收条件:
  30. ```bash
  31. python scripts/react_loop.py --project "D:/my-drama" resume
  32. ```
  33. 开发验证与打包:
  34. ```bash
  35. python -X utf8 -m unittest discover -s tests -v
  36. python scripts/verify_installations.py
  37. python scripts/build_bundle.py
  38. ```
  39. 安装包输出为 `dist/manju-director.zip`,带逐文件哈希清单。
  40. ## 当前范围
  41. 已接入 ViMax 编剧/分镜、take 分镜校验与导出、PyAV 媒体索引/提帧,以及共享核心的 MCP 服务。[组件安装与使用](docs/INTEGRATIONS.md)
  42. 基础循环只用标准库;组件扩展按固定依赖安装。fal 已有任务状态查询适配器,生成提交尚未接入;真实视频质量需要实际素材与视觉评审。程序护栏约束该入口内的调用,视觉观察的真实性由宿主负责。
  43. [运行验证](docs/REACT_RUNTIME_VALIDATION.md) · [研究接入设计](docs/REACT_INTEGRATION_REVIEW.md) · [模型配置](references/model-adapters/README.md)
  44. [MIT License](LICENSE) © 2026 liangie7420

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.