claude-skills
About
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.
Documentation
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
Quick Install
/plugin add https://github.com/2025Emma/vibe-coding-cn/tree/main/claude-skillsCopy and paste this command in Claude Code to install this skill
GitHub 仓库
Related Skills
sglang
MetaSGLang is a high-performance LLM serving framework that specializes in fast, structured generation for JSON, regex, and agentic workflows using its RadixAttention prefix caching. It delivers significantly faster inference, especially for tasks with repeated prefixes, making it ideal for complex, structured outputs and multi-turn conversations. Choose SGLang over alternatives like vLLM when you need constrained decoding or are building applications with extensive prefix sharing.
evaluating-llms-harness
TestingThis Claude Skill runs the lm-evaluation-harness to benchmark LLMs across 60+ standardized academic tasks like MMLU and GSM8K. It's designed for developers to compare model quality, track training progress, or report academic results. The tool supports various backends including HuggingFace and vLLM models.
llamaguard
OtherLlamaGuard is Meta's 7-8B parameter model for moderating LLM inputs and outputs across six safety categories like violence and hate speech. It offers 94-95% accuracy and can be deployed using vLLM, Hugging Face, or Amazon SageMaker. Use this skill to easily integrate content filtering and safety guardrails into your AI applications.
langchain
MetaLangChain is a framework for building LLM applications using agents, chains, and RAG pipelines. It supports multiple LLM providers, offers 500+ integrations, and includes features like tool calling and memory management. Use it for rapid prototyping and deploying production systems like chatbots, autonomous agents, and question-answering services.
