skills-development（技能发展）

Skills Development

Create skills that follow the Agent Skills specification—an open format supported by Claude Code, Cursor, VS Code, GitHub, and other agent products.

Workflow

1. Discovery — Understand what the skill should do
2. Archetype Selection — Choose the best pattern
3. Initialization — Create skill structure from template
4. Customization — Tailor to specific needs
5. Validation — Verify quality before committing

Phase 1: Discovery

Ask about the skill:

• What problem does this skill solve?
• What are the main capabilities?
• What triggers should invoke it? (phrases users would say)
• Where should it live? (personal ~/.claude/skills/, project .claude/skills/, or plugin)

Phase 2: Archetype Selection

Templates in templates/skill-archetypes/.

Phase 3: Initialization

Run the init script:

bun run ${CLAUDE_PLUGIN_ROOT}/skills/skills-development/scripts/init-skill.ts <skill-name> <output-dir> --template <archetype>

Output directories:

• Personal: ~/.claude/skills/<skill-name>/
• Project: .claude/skills/<skill-name>/
• Plugin: <plugin-dir>/skills/<skill-name>/

Phase 4: Customization

Directory Structure

skill-name/
├── SKILL.md           # Required: instructions + metadata
├── scripts/           # Optional: executable code
├── references/        # Optional: documentation
└── assets/            # Optional: templates, resources

Frontmatter

---
name: skill-name
description: What it does and when to use it. Include trigger keywords.
license: Apache-2.0                    # optional
compatibility: Requires git and jq     # optional
metadata:                              # optional
  author: your-org
  version: "1.0"
allowed-tools: Read Grep Glob          # optional, experimental
user-invocable: true                   # optional, makes skill a /command
---

user-invocable

When user-invocable: true, the skill becomes callable as a slash command:

---
name: code-review
description: Reviews code for bugs and best practices...
user-invocable: true
---

Users can then invoke with /code-review instead of waiting for auto-activation. The skill content becomes the command prompt.

Custom Frontmatter

Custom fields must be nested under metadata:

---
name: my-skill
description: ...
metadata:
  author: your-org
  version: "1.0"
  category: development
  tags: [typescript, testing]
---

Top-level custom fields are not allowed and may cause parsing errors.

Description Formula

[WHAT] + [WHEN] + [TRIGGERS]

description: Extracts text and tables from PDF files, fills forms, merges documents. Use when working with PDF files or when the user mentions PDFs, forms, or document extraction.

Checklist:

• Explains WHAT (capabilities)
• States WHEN (trigger conditions)
• Includes 3-5 trigger KEYWORDS
• Uses third-person voice
• Under 200 words

Per-Archetype Customization

api-wrapper:

• Update scripts/client.ts with API endpoints
• Add auth instructions
• Create references/endpoints.md if API is large

document-processor:

• Update scripts/process.ts with format logic
• Add library dependencies
• Document supported operations

dev-workflow:

• Update scripts/run.ts with commands
• Add safety checks for destructive operations
• Document available commands

research-synthesizer:

• Define source priority order
• Specify output format
• Set citation style

Phase 5: Validation

Validate before committing. Use the CLI or run checks manually:

skills-ref validate ./my-skill

Validation Checklist

A. YAML Frontmatter

• Opens with --- on line 1, closes with ---
• name and description present (required)
• Uses spaces, not tabs
• Special characters quoted

B. Naming

• Lowercase, numbers, hyphens only (1-64 chars)
• Matches parent directory name
• No --, leading/trailing hyphens
• No anthropic or claude in name

C. Description Quality

• WHAT: Explains capabilities
• WHEN: States "Use when..." conditions
• TRIGGERS: 3-5 keywords users would say
• Third-person voice (not "I can" or "you can")

D. Structure

• SKILL.md under 500 lines
• All referenced files exist
• No TODO/placeholder markers
• Progressive disclosure (details in references/)

Review Checks

After validation passes, review for quality:

Report Format

# Skill Check: {skill-name}

**Status**: PASS | WARNINGS | FAIL
**Issues**: {critical} critical, {warnings} warnings

## Critical (must fix)
1. {issue with fix}

## Warnings (should fix)
1. {issue with fix}

## Strengths
- {what's done well}

Common Fixes

Vague description:

# Before
description: Helps with PDF files

# After
description: Extracts text and tables from PDF files, fills forms, merges documents. Use when working with PDFs, forms, or document extraction.

Wrong voice:

# Before
description: I can help you process data files

# After
description: Processes data files and converts between formats (CSV, JSON, XML). Use when working with data files or format conversion.

Priority Levels

• Critical: Invalid YAML, missing required fields, broken references
• Important: Vague descriptions, wrong voice, missing triggers
• Nice-to-have: Additional examples, better organization

Core Principles

Concise is key

Context window is shared. Only include what the agent doesn't already know. Challenge each paragraph—does it justify its token cost?

Third-person descriptions

Descriptions inject into system prompt:

• ✓ "Extracts text from PDFs"
• ✗ "I can help you extract text from PDFs"

Progressive disclosure

Keep SKILL.md under 500 lines. Move details to:

• references/ - Detailed docs, API references
• scripts/ - Executable utilities (code never enters context)
• assets/ - Templates, data files

Token loading:

1. Metadata (~100 tokens): name + description at startup
2. Instructions (<5000 tokens): SKILL.md body when activated
3. Resources (as needed): files loaded only when referenced

Degrees of freedom

Match instruction specificity to task requirements:

• High freedom (text): Multiple valid approaches, use judgment
• Medium freedom (pseudocode): Preferred pattern with variation allowed
• Low freedom (scripts): Exact sequence required, no deviation

See patterns.md for detailed examples.

Naming Requirements

• Lowercase letters, numbers, hyphens only
• Cannot start/end with hyphen or contain --
• Must match parent directory name
• Cannot contain anthropic or claude

Recommended: Gerund form (processing-pdfs, reviewing-code)

Tool Restrictions

Use allowed-tools to limit capabilities (experimental):

# Read-only
allowed-tools: Read Grep Glob

# With shell access
allowed-tools: Bash(git:*) Bash(jq:*) Read

# Full access (default)
# Omit field entirely

After Creation

1. Validate: Run Phase 5 checklist or skills-ref validate
2. Test: Ask a question that should trigger it
3. Iterate: Fix issues until passing
4. Share: Commit to git or distribute via plugin

Troubleshooting

# Find all skills
find ~/.claude/skills .claude/skills -name "SKILL.md" 2>/dev/null

# Check for tab characters (YAML requires spaces)
grep -P "\t" SKILL.md

# Find broken markdown links
grep -oE '\[[^]]+\]\([^)]+\)' SKILL.md | while read link; do
  file=$(echo "$link" | sed 's/.*(\(.*\))/\1/')
  [ ! -f "$file" ] && echo "Missing: $file"
done

References

• patterns.md - Degrees of freedom, script design, variant organization, writing patterns
• best-practices.md - Community patterns, testing strategies, advanced techniques
• quick-reference.md - Fast checklist and one-liners
• implementations.md - Per-tool storage paths
• invocations.md - How tools activate skills
• compatibility.md - Path compatibility matrix

Platform-Specific Implementation

Each tool has specific implementation details beyond the common spec:

• Claude Code: claude.md - Tool restrictions syntax, --debug testing, troubleshooting, command/hook integration
• Codex CLI: codex.md - Discovery paths, $skill-name invocation, AGENTS.md relationship, feature flags, built-in skills

See implementations.md for storage paths and platform-specific notes.

External Resources

• Agent Skills Specification
• Best Practices Guide
• skills-ref Validation Library