idea-team

106 Cross-Runtime Skills | 7 Claude Code Agents | One Command Install

审核状态：已审核Quality 80Security 92

复制安装命令

用 Codex 或 Claude 安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它先审查 Skill 页面再帮你安装。

复制前请先查看来源、License 和安全提示。

项目 README

来源文件：README.md

抓取于 2026年9月6日

Spellbook

106 Cross-Runtime Skills | 7 Claude Code Agents | One Command Install

A cross-runtime skill library for Claude Code, Codex, and multi-agent workflows.

Quick Start • Runtime Targets • Pick a Workflow • Skills • Agents • Changelog • Release Status • Contributing • 中文

Rename notice: Spellbook was formerly Claude Arsenal. Claude Code remains a first-class target; the new name reflects the broader roadmap for Claude Code, Codex, and cross-runtime agent skills. See the migration note for details.

Quick Start

Start with one job-shaped workflow. The maintained skills CLI lets you choose the supported coding agents during installation and installs only these four skills:

npx skills add majiayu000/spellbook --skill frontend-design --skill app-ui-design --skill ui-design-system --skill figma-to-react

Use npx skills add majiayu000/spellbook --list to inspect the catalog before installing. See Pick a Workflow for four other focused starting points.

Advanced Cross-Runtime Installer

install.sh remains available when you want explicit Claude Code/Codex target paths or need to install the repository's Claude Code agents as well as skills.

# Install all skills and supported agents into both maintained runtimes
curl -fsSL https://raw.githubusercontent.com/majiayu000/spellbook/main/install.sh | bash -s -- --target all

# Or clone the repository and select skills explicitly
git clone https://github.com/majiayu000/spellbook.git
cd spellbook
./install.sh --target all --skills typescript-project,python-project,devops-excellence

Verify Installation

Claude Code: type / to see your installed skills.
Codex: restart Codex so it reloads ~/.agents/skills.

Runtime Targets

Spellbook keeps the skill source in one place and installs it into the runtime you use.

Target	Installed To	Status
Claude Code	`~/.claude/skills` plus `~/.claude/agents`	Skills and agents supported
Codex	`~/.agents/skills`	Skills supported; agents skipped
All	Both Claude Code and Codex paths	Recommended for multi-tool users

Claude Code remains a first-class target and search entry. The project was formerly known as Claude Arsenal; the new Spellbook name reflects the broader goal: reusable skills that can travel across coding agents. Older Spellbook versions installed Codex skills under ~/.codex/skills; reinstall with the current installer to use the documented Codex user-level skill path.

Pick a Workflow

Start with a small bundle that matches the job, then add more skills when the workflow sticks.

Workflow	Install	Good for
Frontend and UI	`npx skills add majiayu000/spellbook --skill frontend-design --skill app-ui-design --skill ui-design-system --skill figma-to-react`	Product UI, landing pages, design systems, Figma handoff
Code quality	`npx skills add majiayu000/spellbook --skill codebase-audit --skill flowguard --skill systematic-debugging --skill review-gate`	Audits, guarded delivery, root-cause debugging, pre-landing review
Ops and release	`npx skills add majiayu000/spellbook --skill release-engineering --skill server-security --skill clash-doctor --skill system-doctor`	Release planning, server hardening, and local or network diagnosis
Product and docs	`npx skills add majiayu000/spellbook --skill product-discovery --skill prd-master --skill technical-spec --skill product-analytics`	Discovery, PRDs, technical specs, metrics plans
Agent workflows	`npx skills add majiayu000/spellbook --skill codex-agent --skill multi-ai-research --skill flowguard --skill vibeguard`	Cross-review, multi-AI research, context handoff, anti-hallucination checks

High-signal individual skills to try first: github-trending, harmonyos-app, app-ui-design, product-discovery, xiaohongshu, codebase-audit, and server-security.

See Showcase for copy-paste prompts and expected outputs. Use the Spellbook Skill Browser for curated first-party skills, or the Claude Skills Registry for broader community discovery. Release history lives in Changelog.

Why Spellbook

Cross-runtime install: one source tree can install into Claude Code and Codex.
Validated registry: every installable skill is checked by python3 scripts/validate_skills.py --check.
Progressive disclosure: larger skills use references/, templates/, scripts/, and eval files instead of one giant prompt.
Practical coverage: engineering, operations, product, UI, content, and agent workflows live in one catalog.

Skills

The generated full skill inventory lives in Skill Registry. Skill layout rules live in Skill Format Policy. Skill authoring quality rules live in Skill Quality Playbook.

Search the Registry

# Free-text query (AND semantics across name, description, category, tags)
python3 scripts/validate_skills.py search rust testing

# Filter by tag
python3 scripts/validate_skills.py search --tag agent

# Restrict to a description language
python3 scripts/validate_skills.py search --language zh deploy

# Machine-readable output
python3 scripts/validate_skills.py search --tag react --json

The tag index lives in registry/tags.json for tooling and dashboards. Curated overrides for skills the keyword heuristic cannot infer live in registry/tag_overrides.yml.

Audit non-blocking skill quality signals:

python3 scripts/audit_skill_quality.py
python3 scripts/audit_skill_quality.py skill-creator

AI & Agent Workflow

Skills for orchestrating, guarding, and maintaining AI agent workflows — the core of Spellbook's cross-runtime mission.

Skill	Description
`multi-model-orchestrator`	Coordinate multi-agent tasks via a centralized handoff document
`flowguard`	Guard long, ambiguous, or stateful agent tasks from drift
`skill-lifeguard`	Add reliable-skill contracts, checkpoints, smoke hooks, and drift signals
`review-gate`	Produce review packs and require human approval before landing agent changes
`skill-audit`	Audit, design, categorize, and measure agent skills
`skill-ecosystem-doctor`	Govern canonical sources, projections, retirement, quarantine, and cross-runtime verification
`threads`	Codex-native subagents and parallel GitHub queue lanes
`codex-fluent`	Codex session hygiene, archive strategy, and handoff discipline
`codex-retrospective`	Codex self-review of recent history to improve behavior
`brainstorming`	Socratic dialogue for design refinement and architecture exploration

See docs/agent-reliability-trio.md for the Reliable Skill + Context Engineering + Review Gate workflow.

Development Architecture

Build production-ready projects with language-specific best practices.

Skill	Language	Key Features
`typescript-project`	TypeScript	ESM, Zod, Biome, Clean Architecture
`python-project`	Python	uv, Pydantic, Ruff, FastAPI
`rust-project`	Rust	Cargo workspace, error handling, async
`golang-web`	Go	Chi/Echo, sqlc, structured logging
`zig-project`	Zig	Build system, memory management
`architecture-foundation`	Cross-language	Runtime, state ownership, adapters, and convergence specs
`elegant-architecture`	Cross-language	Clean architecture with strict 200-line file limits

Product Lifecycle

End-to-end product development from discovery to deployment.

Skill	Phase	What You Get
`product-discovery`	Discovery	JTBD, user interviews, market research
`prd-master`	Definition	PRD writing, user stories, RICE prioritization
`technical-spec`	Design	Design docs, ADR, C4 diagrams
`product-analytics`	Growth	Event tracking, A/B testing, AARRR
`devops-excellence`	Deployment	CI/CD, Docker, Kubernetes, GitOps
`observability-sre`	Operations	Monitoring, logging, tracing, SLO/SLI
`product-manager-toolkit`	Definition	RICE, customer interviews, PRD templates, discovery frameworks

API & Backend

Skill	Description
`api-design`	REST/GraphQL/gRPC patterns, OpenAPI 3.2
`auth-security`	OAuth 2.1, JWT, security best practices
`database-patterns`	PostgreSQL, Redis, migrations, optimization
`codebase-audit`	Deep adaptive repository audit with severity-ranked findings and repair roadmap
`structured-logging-lite`	Centralized logging, field standards, and distributed tracing

Development Practices

Skill	Description	Origin
`contributor`	End-to-end open source contribution workflow from issue discovery to PR submission	Custom
`repo-agent-context-audit`	Audit and scaffold repo agent context across AGENTS, skills, and specs	Custom
`skill-creator`	Create, improve, and benchmark reusable skills	Custom
`humanizer`	Remove obvious AI writing patterns from user-facing text	External guide + custom adaptation

Delivery Workflow

Disciplined end-to-end delivery: testing, commits, health checks, and contribution flow.

Skill	Description
`app-user-story-qa`	End-to-end app feature inventory, canonical tracker, user-story testing, fixes, and retest loop
`test-driven-development`	Enforce RED-GREEN-REFACTOR TDD discipline
`comprehensive-testing`	Test pyramid, unit/integration/E2E/property testing, framework best practices
`git-commit-smart`	Generate meaningful conventional commit messages from diff
`push-all`	Stage, commit, and push all changes after safety checks
`project-health-auditor`	Codebase health, tech debt, dependency, and project risk analysis
`contribution-architect`	Move from bug fixes to architectural improvements and debt discovery

Cross-Tool Interop

Skills for using multiple coding agents and CLI tools together.

Skill	Description
`codex`	Invoke Codex CLI sessions from another agent workflow
`codex-agent`	Optional second-opinion review, cross-verification, and alternatives through Codex CLI
`sol-luna-router`	Keep GPT-5.6 Sol as commander/reviewer while GPT-5.6 Luna performs bounded implementation
`ask-opencli`	Ask Grok or Gemini through opencli and an existing browser session
`multi-ai-research`	Parallel research across multiple AI tools and internal agents

UI/UX & Design

Skill	Description
`app-ui-design`	iOS/Android UI design, Material Design 3, HIG
`product-ux-expert`	UX evaluation, heuristics, accessibility
`frontend-design`	Web frontend design patterns
`ui-designer`	Extract design systems from UI screenshots and references
`ui-design-system`	Design system toolkit and design-dev handoff support
`web-artifacts-builder`	Claude.ai HTML artifacts
`react-best-practices`	React and Next.js performance patterns distilled from Vercel guidance
`react-hooks-best-practices`	React hooks, effects, refs, and component design patterns
`slides`	Speech-friendly slide deck and background slide generation
`ui-ux-pro-max`	Compact UI/UX tables for product patterns, landing pages, charts, and 9 stacks
`figma-to-code`	Figma designs to production React/Next.js with TypeScript and Tailwind
`css-debug`	Diagnose CSS/layout issues, Tailwind conflicts, z-index stacking
`playwright-automation`	Browser automation and testing with Playwright

Tooling & Automation

Skill	Description
`web-asset-generator`	Favicons, app icons, OG images
`github-trending`	GitHub trending analysis
`vibeguard`	Task contracts, finding scoring, and lightweight anti-hallucination reviews
`clash-doctor`	Clash proxy & network diagnostics
`clash-routes`	Inspect active proxy routes for specific processes via Mihomo API
`optimize-network`	Safe local network speed, latency, DNS, Wi-Fi, and bufferbloat diagnostics with VPN/proxy guardrails
`disk-cleaner`	Scan and reclaim disk space with interactive cleanup guidance
`system-doctor`	Diagnose CPU, memory, and process-level system slowdowns
`codex-log-guard`	Diagnose and mitigate excessive Codex local SQLite diagnostic log writes
`server-security`	Audit and harden Linux server SSH, firewall, and exposed services
`cliproxy-newapi-stack`	Add a loopback-first NewAPI metering layer to an independently verified CLIProxyAPI upstream

Operations & Deploy

Deploy models and diagnose local and remote environments.

Skill	Description
`gemma4-local-deploy`	Deploy Gemma 4 12B locally on Mac/Apple Silicon via llama.cpp or Ollama
`gpu-use`	Inspect remote server GPU usage (per-card VRAM, processes, containers)
`rustdesk-doctor`	Diagnose RustDesk connection issues
`vscode-doctor`	Diagnose slow or freezing VS Code-compatible editors

Content & Social Media

Skill	Description
`xiaohongshu`	Xiaohongshu content creation & publishing
`trip-planner`	Travel itinerary planning
`weekly`	Weekly report from Git, Claude Code, and Codex sessions
`xiaohongshu-netfeel-guardian`	Remove translation-tone from Claude's Chinese content for native readability

Mobile & Cross-Platform

Skill	Description
`harmonyos-app`	HarmonyOS with ArkTS, ArkUI, Stage Model

Rust Specific

Skill	Description
`rust-best-practices`	Microsoft Rust guidelines, error handling

Agents

Specialized agents for complex tasks.

Agent	Expertise	Use Case
`tech-lead-orchestrator`	Coordination	Multi-step tasks, delegation
`code-archaeologist`	Exploration	Legacy codebase documentation
`backend-typescript-architect`	Architecture	Bun/Node.js, API design
`senior-code-reviewer`	Review	Security, performance, architecture
`kubernetes-specialist`	Infrastructure	K8s, Helm, GitOps
`security-auditor`	Security	OWASP Top 10, SAST
`opensource-contributor`	Contribution	Open source workflow

Plugins

Spellbook is also a Claude Code plugin marketplace. Install the repo as a marketplace, then install plugins from it:

/plugin marketplace add majiayu000/spellbook
/plugin install idea-coach
/plugin install rust-dev

Plugin	Description
`idea-coach`	Opinionated product coach (idea -> PRD -> clickable HTML prototype) + multi-role idea group chat; plugin commands are `/idea-coach:idea` and `/idea-coach:idea-team`
`rust-dev`	Rust best practices, code review, performance, and async patterns

Plugin skills are packaged copies of catalog skills; the catalog (installed by install.sh) remains the cross-runtime source of truth.

Skill Design Philosophy

Every skill in Spellbook follows these principles:

Hard Rules - Mandatory constraints with FORBIDDEN / REQUIRED markers
Practical Examples - Real code, not just theory
Verification Checklists - Actionable validation steps
Battle-Tested - Used in production environments

Documentation

Document	Description
Changelog	Release history and current release status
Installation Guide	Detailed setup instructions
Runtime Targets	Claude Code and Codex installation targets
Showcase	Copy-paste workflow demos
Spellbook Operating Contract	Agent behavior rules for autonomy, escalation, pushback, feedback loops, and done-when checks
Skill Format Policy	Directory vs file skill layout rules
Skill Quality Playbook	Trigger descriptions, gotchas, progressive disclosure, and verification
Skill Testing Guide	How to validate skills work
Creating Plugins	Build your own skills
Product Lifecycle (EN)	Full lifecycle coverage
Product Lifecycle (中文)	产品生命周期覆盖

Release Status

Spellbook is in pre-1.0 release-readiness mode. No numbered GitHub release tag has been cut yet; the current install path uses the repository main branch. See Changelog for release history.

Current limitations:

Codex installs skills only; Claude Code agents are skipped for Codex targets.
Some skills depend on external CLIs, accounts, credentials, or platform access that are not bundled by the installer.
The registry validator checks installable skill structure, not every external workflow end to end.

Support paths:

Bugs: open an issue
Feature ideas: open a feature request
Security vulnerabilities: follow Security Policy

Credits

Built on the shoulders of giants:

anthropics/skills - Official Anthropic skills
obra/superpowers - Development methodology
claude-code-plugins-plus - Plugin hub
VoltAgent/awesome-claude-code-subagents - Agent collection

Contributing

Contributions welcome! Please read our Contributing Guide first.

Found a bug? Open an issue
Have a skill idea? Open a feature request
Want to contribute? Submit a PR

The Agent Infra Stack

This project is one layer of an open-source stack for running coding agents (Claude Code, Codex) as serious infrastructure. Every piece works standalone; together they close the loop:

spellbook sits in the Extend layer — the authoring side of the skill story: write once, run on Claude Code and Codex. Discovery and distribution live in claude-skill-registry.

Layer	Project	What it does
Extend	claude-skill-registry	Discover and search community Claude Code skills
Extend	spellbook ◀ you are here	Cross-runtime skills for Claude Code, Codex, and multi-agent workflows
Trust	argus	Static install-time scanner for supply-chain attacks (npm / PyPI / crates.io)
Trust	vibeguard	Rules, hooks, and guards against hallucinated or unverified agent changes
Remember	remem	Local-first persistent memory for Claude Code and Codex sessions
Orchestrate	harness	Rust agent orchestration platform — rules, skills, GC, observability
Route	litellm-rs	High-performance Rust AI gateway — 100+ LLM APIs via OpenAI format
Keep	keepline	Session command center — monitor, recover, never lose agent work

License

MIT License - Use freely in your projects.

If this helps you, consider giving it a ⭐

Made for builders using Claude Code, Codex, and multi-agent workflows

Agent / MCP / Skill 创作

低风险

来源需自行核对维护者身份。
包含脚本或命令调用，安装前请复核。
未检测到明显外部权限要求。
未检测到高风险命令。
扫描发现：0 条。

Codex — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/majiayu000/spellbook.git
将 "skills/idea-team" 文件夹复制到 Codex 的 skills 目录中。
重启 Codex 让新的 skill 生效。

Codex — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Codex 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Codex 让新的 skill 生效。

Claude Code — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/majiayu000/spellbook.git
将 "skills/idea-team" 文件夹复制到 Claude Code 的 skills 目录中。
重启 Claude Code 让新的 skill 生效。

Claude Code — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Claude Code 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Claude Code 让新的 skill 生效。

Cursor — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/majiayu000/spellbook.git
将 "skills/idea-team" 文件夹复制到 Cursor 的 skills 目录中。
重启 Cursor 让新的 skill 生效。

Cursor — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Cursor 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Cursor 让新的 skill 生效。

GitHub Copilot — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/majiayu000/spellbook.git
将 "skills/idea-team" 文件夹复制到 GitHub Copilot 的 skills 目录中。
重启 GitHub Copilot 让新的 skill 生效。

GitHub Copilot — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 GitHub Copilot 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 GitHub Copilot 让新的 skill 生效。

Windsurf — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/majiayu000/spellbook.git
将 "skills/idea-team" 文件夹复制到 Windsurf 的 skills 目录中。
重启 Windsurf 让新的 skill 生效。

Windsurf — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Windsurf 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Windsurf 让新的 skill 生效。

查看 SKILL.md 原文

name: idea-team
description: 想法群聊室主持人 — 把一句话想法丢给多角色 AI 团队（调研员/反方/类比者）做查漏补缺。每个角色有自己的 voice，他们互相 @ 接话；你随时插话。**这是创意扩展工具，不打分、不否决、不堵路**。Use when 用户说"组个团队聊一下"、"开会讨论这个想法"、"找几个角度看看"、"群聊一下 X"、"team review X"，或调用插件命令 `/idea-coach:idea-team`。Do NOT use when 用户已决定做这个想法只要 PRD（用 `idea-to-product`）、或用户要单独某个角色 skill、或用户只想自由 brainstorm（不需要 team 结构）。

idea-team — 想法群聊室

把一句话想法丢给 3 人小组（调研员/反方/类比者），按回合说话，你随时插话。 这是查漏补缺工具，不是 judge — 他们帮你看你没想到的，不堵你的路。

1. 何时触发

插件命令：/idea-coach:idea-team <想法>
Catalog 安装：直接调用 idea-team skill；不承诺插件外的别名
自然语："组个团队聊一下 X"、"开个会讨论 X"、"找几个角度看 X"、"group review X"

2. 何时不要触发

用户要单独某个角色 → 直接调用对应 catalog skill；插件内由 namespaced skill 自动路由
用户已决定做这个想法、要做 PRD → 插件用 /idea-coach:idea，catalog 调用 idea-to-product
用户只要自由 brainstorm 不要团队结构 → 直接对话即可

3. 团队成员（MVP 3 人）

角色	emoji	它做什么	详细 voice 规则
调研员	🔍	拉真实 2026 数据 / 竞品 / 案例	`idea-research/SKILL.md`
反方	⚔️	挑漏洞 / 找隐藏假设 / 给反例	`idea-devils-advocate/SKILL.md`
类比者	🪞	跨界类比 + yes-and 扩展	`idea-analogist/SKILL.md`

角色 SKILL.md 随 skill 安装位置解析，先解析实际位置再 Read：Claude Code 通常是 ~/.claude/skills/<name>/SKILL.md，Codex 通常是 ~/.agents/skills/<name>/SKILL.md。

进入群聊前必须依次 Read 这 3 个 SKILL.md — 加载每个角色的 voice 边界。如果其中某个没装（目录不存在），主持人提醒用户跑 spellbook 仓库根目录的 install.sh，不要替补凑数。

4. 启动流程

抽取 raw_idea
派生 kebab-case slug
检查 .idea-team/<slug>/：若已有 state.json 或 chat.md，先让用户选择 resume 或创建带递增后缀的唯一 slug（如 <slug>-2）；选择前不得覆盖或 append

确认目录唯一后 mkdir -p .idea-team/<slug>/，写 state.json：

{ "slug": "...", "raw_idea": "...", "round": 1, "speakers": [], "round_complete": false, "status": "active", "utterances": [], "report_markdown": null, "created_at": "<ISO 8601>", "updated_at": "<ISO 8601>", "active_roles": ["research", "devils-advocate", "analogist"] }

依次 Read 3 个角色 SKILL.md
进入群聊：主持人开场

5. 群聊机制（核心）

输出格式

每条发言独立一段，但一个回合的 3 个角色必须在同一次 assistant response 中按顺序输出为 3 个独立 block。聊天 runtime 不会在没有新用户消息时自动触发下一次 assistant response，所以禁止把同一回合拆成三个等待触发的回复。

🔍 调研员: <内容，≤ 80 字>。[来源: <来源, 日期>]

⚔️ 反方: @调研员 <挑刺/反例，≤ 80 字>。

🪞 类比者: <类比 + yes-and 扩展，≤ 80 字>。

回合制

一回合 = 3 个角色各说一次，固定顺序：调研员 → 反方 → 类比者。

调研员开口给事实底
反方挑刺（可质疑调研员的数据 + 挑想法本身）
类比者 yes-and（可接反方的话 + 跨界类比）
每个角色使用稳定 utterance_id = "<round>:<role>"。先把 id、role、完整文本写入 state.utterances，同步更新去重后的 speakers；若 speakers 已覆盖所有 active_roles，在同一次 state 写入中设置 round_complete = true。state 必须写到同目录临时文件并原子 rename，作为事实来源。
每次 state 提交成功后，从 state.utterances 完整重建 chat.tmp.md（每条含  marker），校验条目数/marker 唯一后原子 rename 为 chat.md。不得分两次 append marker 和正文。若崩溃，Resume 总是从 canonical utterances 重建完整投影；对旧版仅有 chat 的会话，先一次性导入完整 marker block 到 utterances，再重建。

一回合结束，主持人邀请用户插话：

🎬 主持人: 第 N 回合结束。你想 @ 谁继续？说 "@反方 多说点"，或 "下一轮"，或 "汇总"。

用户插话规则

用户输入	行为
`@角色 ...`	该角色直接回应用户的话（不走回合）
"继续" / "下一轮"	先 `round += 1`、清空 `speakers`、设置 `round_complete = false` 并持久化，再在一次 response 内完成三角色新回合
"够了" / "汇总"	进入汇总环节，写 `team-report.md`
自由文本	主持人判断哪个角色最相关，邀请他接话

角色互相 @（鼓励但不强制）

"🔍 调研员: ... @反方你怎么看这个数据？"
"⚔️ 反方: ... @类比者你那个 X 类比漏了一个前提"
被 @ 的角色下一次发言时 acknowledge 那个 @（点头或反驳）

主持人职责（🎬 = main agent 旁白）

开场介绍到场角色
每回合结束邀请用户插话（不许连跑 2 回合不问用户）
检测到 voice 越界时拉回（例：类比者开始挑刺 → 主持人提醒 "类比者只 yes-and"）
用户说"汇总"时主持总结

Resume 规则

读取 round、speakers、round_complete、utterances 和 chat.md，先从 canonical utterances 原子重建 chat 投影。round_complete 必须从 set(speakers) == set(active_roles) 派生并修复，不能把旧布尔值当唯一事实。若当前回合未完成，只输出缺失角色 block；不得重复已记录角色。若当前回合已完成，先邀请用户选择下一轮或汇总，不得擅自递增回合。

6. 输出物

.idea-team/<slug>/：

文件	内容
`state.json`	{ slug, raw_idea, round, speakers, created_at, active_roles }
`chat.md`	完整对话 transcript（每条发言实时 append，不等汇总）
`team-report.md`	用户说"汇总"后生成的结构化总结

team-report.md 结构

# 💡 idea-team 汇总报告：<raw_idea>

## 🔍 调研员的事实底
- 事实 1（来源）
- 事实 2（来源）

## ⚔️ 反方挖出的漏洞 / 隐藏假设
- 漏洞 1
- 漏洞 2

## 🪞 类比者发现的跨界启示
- 类比 1 → 启示
- 类比 2 → 启示

## 🧩 你可能漏掉的 N 个角度
（基于本次群聊未触及但相关的角度，列 3-5 条）

## 💬 完整对话
见 chat.md

7. 命令变体

/idea-coach:idea-team <想法> — 插件：启动新群聊
/idea-coach:idea-team resume [slug] — 插件：续上次未完成的
/idea-coach:idea-team list — 插件：列 cwd 下所有群聊

Catalog 用户直接调用 idea-team skill 并传相同参数。汇总时先生成完整报告文本，再在同一次原子 state 写入中设置 report_markdown、status = "completed" 和 completed_at，之后从 report_markdown 写 team-report.tmp.md 并原子 rename。 Resume 对 completed 会话先用 canonical report_markdown 修复缺失/截断的报告投影，再只展示报告路径；若迁移旧会话时发现完整 team-report 但 state 仍 active，先导入报告并标 completed，不得重入对话。

8. Red Flags — 主持人在本 skill 里最容易跑偏的偷懒

你脑里的借口	反驳
"每个角色等下一次用户消息再说"	❌ runtime 不会自动续答；同一回合必须在一次 response 内输出 3 个独立角色 block
"调研员说完，反方接的内容不用 @ 他"	鼓励互相 @，让对话有"群聊感"
"类比者觉得反方说错了，挑刺反驳"	❌ 类比者只 yes-and。voice 越界由主持人拉回
"连跑 2 个回合再问用户"	❌ 每回合后必须邀请插话
"调研员凭印象说事实"	❌ 必须 WebSearch + [来源: xxx]
"汇总时只重复 chat.md 内容"	汇总必须有"你可能漏掉的 N 个角度"——超越已聊内容
"用户说 @ 反方但反方刚说过，跳过这次"	被 @ 的角色必须立即回应，无论上一句是不是他说的

9. Important Rules

不打分、不否决、不堵路 — 这是查漏补缺，不是 judge。
voice 边界硬隔离 — 调研员不评判、反方不安慰、类比者不挑刺。越界由主持人拉回。
一回合一条 response、三段角色 block — 固定顺序输出，视觉上分段，不依赖不存在的自动续答。
每回合后邀请插话 — 不许连跑 2 回合不问。
每条 ≤ 80 字 — 群聊节奏 > 长篇大论。
来源必引 — 调研员说事实必须 [来源: xxx, 日期]。
chat.md 实时追写 — 每条发言 append 一行，不等汇总。
角色未安装不替补 — 提醒用户跑 spellbook 根目录 install.sh，不让 main agent 自己 cosplay 缺席的角色。
状态逐条持久化 — 每条发言后更新 speakers；轮次转换先写 round 再输出，Resume 不重复角色。

10. 维护者验证

回合触发、断点续聊和 slug 冲突场景见 evals/evals.json。修改回合或持久化语义时，必须同步更新这些 eval。

发现问题？提交给管理员复核

暂无评论，成为第一个评论者吧！