feat(skill): publish /teamai skill to simplify onboarding and daily usage

背景

Supersedes #517。

#517 提出了三份可粘贴提示词帮助不会 Git 的用户接入 TeamAI，但静态提示词维护成本高、容易与 CLI 命令脱节、用户不知该用哪份。

更好的方式：发布标准 SKILL.md 格式的 teamai skill，上架各主流 agent-skill 市场，用户在任何 AI 工具里搜索安装、交互式引导完成接入。

设计原则

仅手动触发

Skill 必须仅通过 /teamai 手动触发，不能通过提示词自动匹配触发，避免用户正常对话中意外激活。

渐进式披露

用户输入 /teamai 时，不做任何操作，仅展示分场景使用示例：

teamai — Team AI Skills & Rules Sync

Usage examples (copy one to get started):

  🏗️  Admin — set up a new team repo:
      /teamai Help me set up TeamAI for my team from scratch

  🤝  Member — join an existing team:
      /teamai Help me join my team's TeamAI, repo URL is https://...

  🔧  Admin — daily management:
      /teamai I already have TeamAI set up, help me manage it

  💡  Member — contribute skills & learnings:
      /teamai I want to contribute what I learned to my team

用户选择某个场景后，skill 才展开对应的详细引导流程。

场景覆盖

场景	面向	核心流程
安装 & 建库	管理员	环境检测 → 平台选择（GitHub/GitLab/CNB）→ 安装 → teamai init → teamai doctor
成员接入	成员	安装 → 登录 → teamai init <URL> → teamai doctor → 验证 skill 生效
日常管理	管理员	push/pull、邀请成员、doctor 排障、roles/packages/env 管理
做贡献	成员	teamai contribute、编写 skill/learning/doc、teamai push

不做的事

• 不教 Git（用户不需要知道底层是 Git）
• 不用 owner/repo 短地址（必须完整 URL）

兼容性要求

模型兼容

Skill 必须在以下两类模型上都能正常工作：

顶尖模型（复杂指令跟随能力强）：

• Claude Opus
• GPT-6

普惠模型（指令跟随能力较弱，覆盖面广）：

• DeepSeek 4.1 Flash
• GLM 5.4 Flash
• Hunyuan 4 (HY4)

提示词需要足够清晰、结构化，不能依赖模型的高级推理能力。

Agent 兼容性测试

发布前必须在以下 agent 中完成端到端测试：

• Claude Code（CLI + IDE）
• ChatGPT App
• CodeBuddy
• WorkBuddy
• Cursor
• 其他支持 SKILL.md 标准的 agent（OpenCode、Gemini CLI 等，按需）

每个 agent 至少完成一次完整的「安装 → init → doctor 通过」流程。

Agent 特殊注意事项

Agent	注意事项
ChatGPT App	沙箱环境下 teamai init 注入的钩子默认不受信任，需引导用户在设置中手动信任（trust）钩子后才能生效
WorkBuddy	沙箱环境不支持 teamai init 后自动添加钩子，需引导用户手动编辑配置文件来注册钩子

Skill 引导流程中遇到上述 agent 时，必须主动提示用户完成对应的额外步骤。

发布渠道

• GitHub 仓库（teamai-skill 或 teamai-cli 内 skills/ 目录）
• Cursor Marketplace
• VoltAgent/awesome-agent-skills PR
• SkillHub (Tencent Cloud) / SkillHub.cn（腾讯官方）
• ClawHub（OpenClaw 社区）

参考

• #554 — OpenClaw 支持改进
• #517 — 原始提示词 RFC
• PR #545 — teamai members --invite 实现
• SKILL.md 标准

---

Background

Supersedes #517.

#517 proposed three copy-paste prompts for users unfamiliar with Git. Static prompts are hard to maintain, drift from actual CLI commands, and leave users guessing which one to use.

A better approach: publish a standard SKILL.md-format teamai skill to major agent-skill marketplaces so users can search, install, and get interactive guidance in any AI tool.

Design Principles

Manual trigger only

The skill MUST only be triggered via /teamai (or equivalent manual invocation). It must NOT auto-trigger via prompt matching to avoid accidental activation during normal conversation.

Progressive disclosure

When user types /teamai with no arguments, take no action — only display usage examples for each scenario:

teamai — Team AI Skills & Rules Sync

Usage examples (copy one to get started):

  🏗️  Admin — set up a new team repo:
      /teamai Help me set up TeamAI for my team from scratch

  🤝  Member — join an existing team:
      /teamai Help me join my team's TeamAI, repo URL is https://...

  🔧  Admin — daily management:
      /teamai I already have TeamAI set up, help me manage it

  💡  Member — contribute skills & learnings:
      /teamai I want to contribute what I learned to my team

The skill expands into detailed guidance only after user selects a scenario.

Scenario coverage

Scenario	Audience	Core flow
Install & create repo	Admin	Env detection → platform selection (GitHub/GitLab/CNB) → install → teamai init → teamai doctor
Member onboarding	Member	Install → login → teamai init <URL> → teamai doctor → verify skill works
Daily management	Admin	push/pull, invite members, doctor troubleshooting, roles/packages/env
Contribute	Member	teamai contribute, write skill/learning/doc, teamai push

Out of scope

• Don't teach Git (users don't need to know the underlying mechanism)
• Don't use owner/repo short addresses (must use full URL)

Compatibility Requirements

Model compatibility

The skill must work correctly on both tiers:

Frontier models (strong instruction following):

• Claude Opus
• GPT-6

Accessible models (weaker instruction following, wide coverage):

• DeepSeek 4.1 Flash
• GLM 5.4 Flash
• Hunyuan 4 (HY4)

The skill prompt must be clear and structured enough to not rely on advanced reasoning.

Agent compatibility testing

Must complete end-to-end testing before publish:

• Claude Code (CLI + IDE)
• ChatGPT App
• CodeBuddy
• WorkBuddy
• Cursor
• Other SKILL.md-compatible agents as needed (OpenCode, Gemini CLI, etc.)

Each agent must complete at least one full "install → init → doctor pass" flow.

Agent-specific caveats

Agent	Caveat
ChatGPT App	Hooks injected by teamai init are untrusted by default in the sandbox — guide the user to manually trust the hooks in settings before they take effect
WorkBuddy	The sandbox does not support adding hooks automatically after teamai init — guide the user to manually edit the config file to register hooks

The skill's guided flow must proactively prompt the user to complete these extra steps when running on the above agents.

Distribution Channels

• GitHub repo (teamai-skill or skills/ directory within teamai-cli)
• Cursor Marketplace
• VoltAgent/awesome-agent-skills PR
• SkillHub (Tencent Cloud) / SkillHub.cn (Tencent official)
• ClawHub (OpenClaw community)

References

• #554 — OpenClaw support improvements
• #517 — Original prompt RFC
• PR #545 — teamai members --invite implementation
• SKILL.md standard

Start by reviewing the existing CLI flows referenced in the issue, especially teamai init, teamai doctor, teamai contribute, and teamai push, then determine whether the skill belongs in a skills/ directory or a separate teamai-skill repository. Done means a manual /teamai entry point covers the listed scenarios, handles the agent-specific caveats, and passes install → init → doctor testing across the required agents and publication channels.