claude-skills
关于
This meta-skill helps developers create and refine reusable Claude Skills by extracting domain material into structured components like SKILL.md and reference folders. It enables building maintainable skills with reliable activation triggers and clear boundaries. Use it to craft new skills from documentation or refactor existing ones for improved clarity and quality.
快速安装
Claude Code
推荐/plugin add https://github.com/2025Emma/vibe-coding-cngit clone https://github.com/2025Emma/vibe-coding-cn.git ~/.claude/skills/claude-skills在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Claude Skills Meta-Skill
Turn scattered domain material into a Skill that is reusable, maintainable, and reliably activatable:
SKILL.mdas the entrypoint (triggers, constraints, patterns, examples)references/for long-form evidence and navigation- optional
scripts/andassets/for scaffolding and templates
When to Use This Skill
Trigger this meta-skill when you need to:
- Create a new Skill from scratch from docs/specs/repos
- Refactor an existing Skill (too long, unclear, inconsistent, misfires)
- Design reliable activation (frontmatter + triggers + boundaries)
- Extract a clean Quick Reference from large material
- Split long content into navigable
references/ - Add a quality gate and a validator
Not For / Boundaries
This meta-skill is NOT:
- A domain Skill by itself (it builds domain Skills)
- A license to invent external facts (if the material does not prove it, say so and add a verification path)
- A substitute for required inputs (if inputs are missing, ask 1-3 questions before proceeding)
Quick Reference
Deliverables (What You Must Produce)
Your output MUST include:
- A concrete directory layout (typically
skills/<skill-name>/) - An actionable
SKILL.mdwith decidable triggers, boundaries, and reproducible examples - Long-form docs moved to
references/with areferences/index.md - A pre-delivery checklist (Quality Gate)
Recommended Layout (Minimal -> Full)
skill-name/
|-- SKILL.md # Required: entrypoint with YAML frontmatter
|-- references/ # Optional: long-form docs/evidence/index
| `-- index.md # Recommended: navigation index
|-- scripts/ # Optional: helpers/automation
`-- assets/ # Optional: templates/configs/static assets
The truly minimal version is just SKILL.md (you can add references/ later).
YAML Frontmatter (Required)
---
name: skill-name
description: "What it does + when to use (activation triggers)."
---
Frontmatter rules:
nameMUST match^[a-z][a-z0-9-]*$and SHOULD match the directory namedescriptionMUST be decidable (not "helps with X") and include concrete trigger keywords
Minimal SKILL.md Skeleton (Copy/Paste)
---
name: my-skill
description: "[Domain] capability: includes [capability 1], [capability 2]. Use when [decidable triggers]."
---
# my-skill Skill
One sentence that states the boundary and the deliverable.
## When to Use This Skill
Trigger when any of these applies:
- [Trigger 1: concrete task/keyword]
- [Trigger 2]
- [Trigger 3]
## Not For / Boundaries
- What this skill will not do (prevents misfires and over-promising)
- Required inputs; ask 1-3 questions if missing
## Quick Reference
### Common Patterns
**Pattern 1:** one-line explanation
```text
[command/snippet you can paste and run]
Examples
Example 1
- Input:
- Steps:
- Expected output / acceptance:
Example 2
Example 3
References
references/index.md: navigationreferences/...: long-form docs split by topic
Maintenance
- Sources: docs/repos/specs (do not invent)
- Last updated: YYYY-MM-DD
- Known limits: what is explicitly out of scope
### Authoring Rules (Non-negotiable)
1. Quick Reference is for short, directly usable patterns
- Keep it <= 20 patterns when possible.
- Anything that needs paragraphs of explanation goes to `references/`.
2. Activation must be decidable
- Frontmatter `description` should say "what + when" with concrete keywords.
- "When to Use" must list specific tasks/inputs/goals, not vague help text.
- "Not For / Boundaries" is mandatory for reliability.
3. No bluffing on external details
- If the material does not prove it, say so and include a verification path.
### Workflow (Material -> Skill)
Do not skip steps:
1. Scope: write MUST/SHOULD/NEVER (three sentences total is fine)
2. Extract patterns: pick 10-20 high-frequency patterns (commands/snippets/flows)
3. Add examples: >= 3 end-to-end examples (input -> steps -> acceptance)
4. Define boundaries: what is out-of-scope + required inputs
5. Split references: move long text into `references/` + write `references/index.md`
6. Apply the gate: run the checklist and the validator
### Quality Gate (Pre-delivery Checklist)
Minimum checks (see `references/quality-checklist.md` for the full version):
1. `name` matches `^[a-z][a-z0-9-]*$` and matches the directory name
2. `description` states "what + when" with concrete trigger keywords
3. Has "When to Use This Skill" with decidable triggers
4. Has "Not For / Boundaries" to reduce misfires
5. Quick Reference is <= 20 patterns and each is directly usable
6. Has >= 3 reproducible examples
7. Long content is in `references/` and `references/index.md` is navigable
8. Uncertain claims include a verification path (no bluffing)
9. Reads like an operator's manual, not a documentation dump
Validate locally:
```bash
# From repo root (basic validation)
./skills/claude-skills/scripts/validate-skill.sh skills/<skill-name>
# From repo root (strict validation)
./skills/claude-skills/scripts/validate-skill.sh skills/<skill-name> --strict
# From skills/claude-skills/ (basic validation)
./scripts/validate-skill.sh ../<skill-name>
# From skills/claude-skills/ (strict validation)
./scripts/validate-skill.sh ../<skill-name> --strict
Tools & Templates
Generate a new Skill skeleton:
# From repo root (generate into ./skills/)
./skills/claude-skills/scripts/create-skill.sh my-skill --full --output skills
# From skills/claude-skills/ (generate into ../ i.e. ./skills/)
./scripts/create-skill.sh my-skill --full --output ..
# Minimal skeleton
./skills/claude-skills/scripts/create-skill.sh my-skill --minimal --output skills
Templates:
assets/template-minimal.mdassets/template-complete.md
Examples
Example 1: Create a Skill from Docs
- Input: an official doc/spec + 2-3 real code samples + common failure modes
- Steps:
- Run
create-skill.shto scaffoldskills/<skill-name>/ - Write frontmatter
descriptionas "what + when" - Extract 10-20 high-frequency patterns into Quick Reference
- Add >= 3 end-to-end examples with acceptance criteria
- Put long content into
references/and wirereferences/index.md - Run
validate-skill.sh --strictand iterate
- Run
Example 2: Refactor a "Doc Dump" Skill
- Input: an existing
SKILL.mdwith long pasted documentation - Steps:
- Identify which parts are patterns vs. long-form explanation
- Move long-form text into
references/(split by topic) - Rewrite Quick Reference as short copy/paste patterns
- Add or fix Examples until they are reproducible
- Add "Not For / Boundaries" to reduce misfires
Example 3: Validate and Gate a Skill
- Input:
skills/<skill-name>/ - Steps:
- Run
validate-skill.sh(non-strict) to get warnings - Fix frontmatter/name mismatches and missing sections
- Run
validate-skill.sh --strictto enforce the spec - Run the scoring rubric in
references/quality-checklist.mdbefore shipping
- Run
References
Local docs:
references/index.mdreferences/skill-spec.mdreferences/quality-checklist.mdreferences/anti-patterns.mdreferences/README.md(upstream official reference)
External (official):
- https://support.claude.com/en/articles/12512176-what-are-skills
- https://support.claude.com/en/articles/12512180-using-skills-in-claude
- https://support.claude.com/en/articles/12512198-creating-custom-skills
- https://docs.claude.com/en/api/skills-guide
Maintenance
- Sources: local spec files in
skills/claude-skills/references/+ upstream official docs inreferences/README.md - Last updated: 2025-12-14
- Known limits:
validate-skill.shis heuristic; strict mode assumes the recommended section headings
GitHub 仓库
相关推荐技能
creating-opencode-plugins
元该Skill为开发者创建OpenCode插件提供指导,涵盖命令、文件、LSP等25+种事件类型。它详细说明了插件结构、事件API规范及JavaScript/TypeScript实现模式,帮助开发者构建事件驱动的模块。适用于需要拦截操作、扩展功能或自定义AI助手行为的插件开发场景。
sglang
元SGLang是一个专为LLM设计的高性能推理框架,特别适用于需要结构化输出的场景。它通过RadixAttention前缀缓存技术,在处理JSON、正则表达式、工具调用等具有重复前缀的复杂工作流时,能实现极速生成。如果你正在构建智能体或多轮对话系统,并追求远超vLLM的推理性能,SGLang是理想选择。
evaluating-llms-harness
测试该Skill通过60+个学术基准测试(如MMLU、GSM8K等)评估大语言模型质量,适用于模型对比、学术研究及训练进度追踪。它支持HuggingFace、vLLM和API接口,被EleutherAI等行业领先机构广泛采用。开发者可通过简单命令行快速对模型进行多任务批量评估。
langchain
元LangChain是一个用于构建LLM应用程序的框架,支持智能体、链和RAG应用开发。它提供多模型提供商支持、500+工具集成、记忆管理和向量检索等核心功能。开发者可用它快速构建聊天机器人、问答系统和自主代理,适用于从原型验证到生产部署的全流程。
