agents-skills（Agent 技巧）

Agent Skills - Meta Reference

This skill provides the definitive reference for creating, organizing, and maintaining agent skills. Use this when building new skills or improving existing skill architecture.

Quick Reference

Cross-Platform: Agent Skills are designed to be portable across runtimes (Claude Code, Codex CLI, Gemini CLI, VS Code Copilot).

Skill Structure

skills/
	└── skill-name/
	    ├── SKILL.md           # Main reference (required)
	    ├── scripts/           # Executable code (Python/Bash) - prefer running; read only to patch/review
	    │   └── validate.py
	    ├── references/        # Documentation loaded into context on-demand
	    │   ├── patterns.md
	    │   └── examples.md
    ├── assets/            # Files for OUTPUT (templates, icons, fonts)
    │   └── template.html
    └── data/
        └── sources.json   # External references

Directory purposes:

• scripts/ - Executable code; prefer running scripts and consuming output (read only when you need to modify/review)
• references/ - Documentation loaded into context when the agent needs it
• assets/ - Files used in generated output (not loaded into context)

Notes:

• Prefer executing scripts instead of pasting large code into context; read scripts only when you need to modify or review them.

SKILL.md Template

---
name: skill-name
description: One-line description of what this skill provides and when to use it (include trigger keywords here)
---

# Skill Name - Quick Reference

Brief overview of the skill's purpose and value.

## Quick Reference

| Task | Tool/Method | When to Use |
|------|-------------|-------------|
| Task 1 | Tool A | Context for usage |

## Scope (Optional)

Keep this brief; the primary trigger mechanism is the frontmatter `description`.

## Core Concepts

### Concept 1

Explanation with code example:

\`\`\`language
code example
\`\`\`

### Concept 2

Additional patterns...

## Navigation

**Resources**
- [references/skill-patterns.md](references/skill-patterns.md) - Common patterns
- [references/skill-validation.md](references/skill-validation.md) - Validation criteria

**Related Skills**
- [../agents-subagents/SKILL.md](../agents-subagents/SKILL.md) - Agent creation

Progressive Disclosure

Skills use progressive disclosure to optimize token usage:

Pattern: SKILL.md provides overview -> Resources load only when needed

Limits: Keep SKILL.md under 500 lines (<5K tokens)

When to Split Content

Frontmatter Specification

See references/frontmatter-reference.md for full specification, examples, and string substitutions.

Name rules:

• Use kebab-case: ai-llm, not AI_LLM_Engineering
• Match folder name exactly
• Be specific: software-backend not backend
• No claude or anthropic in names (reserved words)
• No XML angle brackets (< >) in any frontmatter value

Description rules (PRIMARY TRIGGER):

• The runtime uses descriptions to decide which skills to auto-invoke
• Format: [What it does]. Use when [trigger phrases].
• Target ~150 chars when library has 50+ skills (budget: 2% of context window, ~16K chars shared across ALL skill descriptions)
• Include key technologies/concepts as trigger keywords
• Single-line YAML only — no multiline >-

String substitutions (usable in skill body):

• $ARGUMENTS / $ARGUMENTS[N] / $N — user-provided arguments from /skill-name arg1 arg2
• ${CLAUDE_SESSION_ID} — current session identifier
• ${CLAUDE_SKILL_DIR} — absolute path to this skill's directory

Dynamic context injection: Use `!command ` in the skill body to inject command output at load time.

Skill Categories

sources.json Schema

{
  "metadata": {
    "title": "Skill Name - Sources",
    "description": "Brief description",
    "last_updated": "YYYY-MM-DD",
    "skill": "skill-name"
  },
  "category_name": [
    {
      "name": "Resource Name",
      "url": "https://example.com/docs",
      "description": "What this covers",
      "add_as_web_search": true
    }
  ]
}

Categories should group logically:

• official_documentation
• tutorials
• community_resources
• tools_and_libraries

Quality Checklist

SKILL VALIDATION CHECKLIST

Frontmatter:
[ ] name matches folder name (kebab-case)
[ ] description is concise and actionable

Structure:
[ ] SKILL.md under 500 lines (split into references/ if needed)
[ ] references/ for detailed content
[ ] data/sources.json with curated links

Content:
[ ] Quick reference table at top
[ ] Frontmatter `description` includes trigger keywords
[ ] Code examples are copy-paste ready
[ ] Related skills linked at bottom

Quality:
[ ] >40% operational content (code, tables, checklists)
[ ] <50% prose paragraphs
[ ] All URLs are live (no 404s)
[ ] Sources updated within 6 months

Multi-Tech vs Single-Tech Skills

Single-Tech Skill

software-backend/
├── SKILL.md              # Node.js focus
└── references/
    └── nodejs-patterns.md

Multi-Tech Skill

software-backend/
├── SKILL.md              # Overview + decision tree
├── references/
│   ├── nodejs-patterns.md
│   ├── go-patterns.md
│   ├── rust-patterns.md
│   └── python-patterns.md
└── assets/
    ├── nodejs/
    ├── go/
    ├── rust/
    └── python/

Navigation

Resources

• references/anthropic-skills-guide.md - Anthropic's official skill-building guide (distilled)
• references/frontmatter-reference.md - Complete frontmatter field specification
• references/skill-patterns.md - Common skill patterns
• references/skill-validation.md - Validation criteria
• data/sources.json - Official documentation links

Related Skills

• ../agents-subagents/SKILL.md - Agent creation
• ../agents-hooks/SKILL.md - Hook automation
• ../agents-mcp/SKILL.md - MCP server integration

Fact-Checking

• Use web search/web fetch to verify current external facts, versions, pricing, deadlines, regulations, or platform behavior before final answers.
• Prefer primary sources; report source links and dates for volatile information.
• If web access is unavailable, state the limitation and mark guidance as unverified.