Meta-Skill Guide: Creating Claude Code Custom Skills

Quick Reference

Aspect	Constraint	Recommendation
SKILL.md size	No hard limit	Keep under 500 lines
Skill name	Max 64 chars	Lowercase, hyphens, gerund form (processing-pdfs)
Description	Max 1024 chars	WHAT it does + WHEN to use (third person)
Reference depth	One level deep	SKILL.md -> File.md only
File paths	Forward slashes	Always scripts/helper.py
Subagent context	Completely isolated	NO conversation history, NO memory

---

Architecture: Three-Level Progressive Loading

Level 1: Metadata (~100 tokens/skill)

---
name: skill-name
description: Brief description with usage triggers
---

• Loaded at Claude Code startup
• Critical for skill discovery

Level 2: Instructions (<5,000 tokens)

• Main SKILL.md body content
• Loaded when user request matches description
• Optimal: Under 500 lines

Level 3: Resources (Unlimited)

• Additional markdown files, scripts, templates
• Loaded as needed via bash navigation
• Scripts execute without loading code into context

---

File Structure

skill-name/
├── SKILL.md              # Main instructions (mandatory)
├── reference/            # Additional documentation
│   ├── advanced.md
│   └── api_reference.md
├── scripts/              # Executable code
│   └── validator.py
└── templates/            # Templates
    └── template.ext

---

YAML Frontmatter

Required:

---
name: skill-name
description: What skill does and when to use it. Include triggers and key terms.
---

Optional (Claude Code):

allowed-tools: ["Bash", "Read", "Write", "Edit"]
model: claude-sonnet-4-5-20250929

Naming Rules:

• Max 64 chars, lowercase letters/numbers/hyphens
• Use gerund form: processing-pdfs (not process-pdf)
• Avoid vague names: helper, utils, documents

---

Skill Levels

Level	Lines	Characteristics	Template
Simple	100-300	Single task, no scripts	reference/templates/simple.md
Moderate	300-500	Multiple ops, validation scripts	reference/templates/moderate.md
Advanced	500+ distributed	Multi-phase, extensive refs	reference/templates/advanced.md

---

Best Practices Summary

Top 10 DOs

1. Keep SKILL.md under 500 lines
2. Write descriptions in third person with triggers
3. Use consistent terminology throughout
4. Provide concrete examples with input/output pairs
5. Implement validation loops (Create -> Validate -> Fix)
6. Structure long files with table of contents
7. Use checklists for complex tasks
8. Solve problems in scripts, don't punt to Claude
9. Justify all constants and magic numbers
10. List required packages explicitly

Top 10 DON'Ts

1. Reference time-sensitive information
2. Use Windows-style paths (scripts\helper.py)
3. Assume tools are pre-installed
4. Use vague MCP tool references
5. Provide too many equal options (decision paralysis)
6. Create deeply nested references (one level max)
7. Skip validation for quality tasks
8. Mix inconsistent terminology
9. Leave constants unexplained
10. Assume subagent context from conversation

Full checklist: reference/best-practices.md

---

Subagent Integration Summary

When to Delegate

Task Type	Recommendation
Simple, <10 min	Direct in skill
Research-heavy	Delegate to context-collector
Multi-stage	Orchestrate specialists
Language-specific	Language expert subagents
Quality assurance	code-reviewer, test-automator

Key Constraints

• NO conversation history access - each invocation is fresh
• NO nested delegation - main agent orchestrates all
• File artifacts for state transfer - save results to files
• Explicit output format - prevent context overload

Orchestration Patterns

1. Parallel - Independent tasks, single message, multiple Task calls
2. Sequential - Dependent tasks, file artifacts between steps
3. Hub-and-Spoke - Implementation + multiple reviewers
4. Iterative - Architect -> Reviewer cycles

Full guide: reference/subagent-patterns.md

---

Domain Patterns Summary

Domain	Key Pattern	Example
Creative	Philosophy-first, reductive mastery	algorithmic-art
Document	Validation loops, zero-error mandate	xlsx, pdf
Development	Phase-based workflow, quality gates	mcp-builder
Enterprise	Template routing, clarification protocols	internal-comms
Workflow	Slash commands, automation	worktree-manager

Full patterns: reference/domain-patterns.md

---

Quality & Testing Summary

Build Evaluations First

1. Run Claude WITHOUT skill on real tasks
2. Create 3+ test scenarios
3. Write minimal instructions to pass
4. Iterate based on real behavior

Two-Claude Method

• Claude A: Skill author
• Claude B: Skill tester
• Iterate based on where B struggles

Cross-Model Testing

• Haiku (cost optimization)
• Sonnet (standard)
• Opus (complex tasks)

Full guide: reference/evaluation.md

---

Key Takeaways

Architecture

• Three-level progressive loading: metadata -> instructions -> resources
• Filesystem-based: bash navigation, scripts execute without loading
• One level deep references only

Structure

• SKILL.md under 500 lines
• YAML: name (64 chars) + description (1024 chars, third person)
• Forward slashes always, gerund naming

Subagents

• Context isolation: NO history, NO memory
• Main agent orchestrates all delegation
• File artifacts for state transfer
• 90% improvement with parallel specialists

Distribution

• Skills don't sync across surfaces
• Claude Code: ~/.claude/skills/ or .claude/skills/
• Share via git for teams

---

External References

Official Docs:

• Skills Overview: https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview
• Best Practices: https://docs.claude.com/en/docs/agents-and-tools/agent-skills/best-practices
• Claude Code Skills: https://docs.claude.com/en/docs/claude-code/skills

GitHub Examples:

• Anthropic Skills: https://github.com/anthropics/skills
• Claude Cookbooks: https://github.com/anthropics/claude-cookbooks/tree/main/skills

Full links: reference/links.md

---

Version

v1.2.0 (December 2025) - Refactored to progressive disclosure pattern