Structure

• SKILL.md (required) - Instructions
• references/ (optional) - Detailed docs
• scripts/ (optional) - Executable code

Frontmatter: name (lowercase-hyphens), description (WHAT + WHEN)

Progressive Disclosure

Metadata (~100 tokens) loads at startup. SKILL.md (<5000 tokens) loads on activation. References load only when explicitly linked (not on activation). Keep SKILL.md lean.

Reference Loading

References are JIT (just-in-time) loaded:

• Only files explicitly linked via [text](references/file.md) load
• Link to files, not folders - [Recipes](references/recipes/README.md) not [Recipes](references/recipes/)
• Each file loads in full (not sections)
• No caching between requests - write self-contained files
• Use recipes/services patterns for multi-option skills

See REFERENCE-LOADING.md for details.

Validation

# Run from the scripts directory
cd scripts
npm run references              # Validate all skill links
npm run tokens -- check         # Check token limits

Integrity Checks

When reviewing or authoring skills, verify:

1. No broken links - All referenced files exist
2. No orphaned references - All reference files are linked
3. Token budgets - References under 1000 tokens (split if exceeded)
4. No duplicates - Consolidate repeated content
5. No out-of-place guidance - Service-specific content belongs in service-specific references

See Validation for detailed procedures.

Reference Documentation

• Guidelines - Detailed writing guidelines
• Token Budgets - Limits and splitting guidance
• Reference Loading - How references load
• Checklist - Pre-submission checklist
• Validation - Link and reference validation
• agentskills.io/specification - Official spec

markdown-token-optimizer

分析并精简 Markdown 内容，降低 token 消耗并提升 AI 处理效率。

creating-skills

帮助你规划并创建新的自定义技能，生成结构、说明文档与可选脚本