| name | skill-dev |
| description | Skill 全生命周期管理：创建 → 反思优化 → 评测 → 成熟度判断 → 发布到市场 → 检索多版本 → 选择/安装 → 融合迭代 → 卸载。触发场景：(1) 用户要求创建/修改 skill (2) 发现可提取为 skill 的重复模式 (3) skill 执行出错或用户纠正后需要反思改进 (4) 用户要求发布/搜索/安装/合并社区 skill (5) 反思后自动检查成熟度并建议发布 |
Skill Dev
Full lifecycle management for AI agent skills: create, reflect, evaluate, publish, search, install, merge, review, and uninstall.
Core Principles
Concise is Key
The context window is a shared resource. Only add context Claude doesn't already have. Challenge each piece: "Does Claude really need this?" Prefer concise examples over verbose explanations.
Degrees of Freedom
Match specificity to the task's fragility:
- High freedom (text instructions): Multiple approaches valid, context-dependent decisions
- Medium freedom (pseudocode/scripts with params): Preferred pattern exists, some variation OK
- Low freedom (specific scripts): Operations fragile, consistency critical, exact sequence required
Routing
Local lifecycle (no registry needed):
- Creating/structuring a skill (directory layout, SKILL.md format, progressive disclosure, verification) → 读取
references/structure.md
- Reflecting after skill failure (trigger signals, reflect process, impact scan, escalation) → 读取
references/reflect-mode.md
- 评测 skill prompt (eval、跑回归、检查 prompt 改动效果) → 读取
references/eval-mode.md
- Checking skill maturity (after reflect, after successful runs, "成熟了吗", "该发布了吗") → 读取
references/maturity.md
Registry lifecycle (public registry built-in, works out of the box):
- Publishing a skill ("publish", "发布 skill", "开源这个 skill") → 读取
references/publish.md
- Searching/installing ("search skill", "有没有XX的skill", "安装 skill") → 读取
references/search.md
- Reviewing a skill ("review", "评价 skill", "打分") → use
scripts/review.py
- Merging skill variants ("merge", "合并版本", "融合") → 读取
references/merge.md
- Uninstalling a skill ("uninstall", "删除 skill", "卸载") → use
scripts/uninstall.py --name <skill> --yes
When NOT to Create a Skill
Don't build for hypothetical future needs. Skip if ANY apply:
- Used only once — just do it inline
- A one-line CLAUDE.md rule covers it — just edit CLAUDE.md directly instead
- No reusable script AND no non-obvious knowledge — Claude already knows how
- An existing skill handles 80%+ of the use case — extend it instead
Script Design
Tool design matters more than prompt design. When a skill has scripts/, invest in quality:
- Token-light output: Print only what the caller needs.
--verbose for debugging only.
- Greppable errors: All errors start with
ERROR: with key details on same line.
- Self-documenting: Support
--help with one-liner description and parameter list.
- Clear parameter names: Use intuitive names (
--document-id, not --did).
- Absolute paths: Accept and output absolute paths.
- Exit codes: 0 = success, non-zero = failure.
- Design for agents, not humans: Output structured data, not formatted text.
- Progressive disclosure: Truncated output must include total data size and how to see more. JSON: add
total/has_more/page_token. Text: append (N chars total) + stderr HINT: with continuation command.
Writing Guidelines
- Do include: non-obvious procedures, domain specifics, gotchas from real failures
- Don't include: things Claude already knows, verbose explanations, auxiliary docs
- Keep SKILL.md ≤150 lines (routing layer); move scenario details to references/
- Challenge each line: "Would removing this cause Claude to make mistakes?" If not, cut it.
- Prefer examples over explanations: One concrete pair teaches more than a paragraph
