"Where Natural Language becomes Turing Complete"
Overview • Usage • How It Works •
Skill Master is an intelligent skill orchestrator for Claude Code that automatically searches, creates, executes, and improves skills based on actual usage.
- Search - Finds existing skills (local, GitHub, web)
- Create - Generates new skills through deep research when none exist
- Execute - Invokes the skill to complete your task
- Review - Compares execution against skill instructions (empirical, not theoretical)
- Improve - Updates skills based on actual divergences
The review is self-limiting: if a skill executes perfectly, no improvement is needed. Skills converge toward optimal instructions through real usage.
Simply invoke Skill Master with your task:
/skill-master Create a business plan for an electric motorcycle startup
Or let Claude Code auto-detect based on context:
Invoke skill-master to create a comprehensive market analysis for renewable energy in Europe
Skill Master will:
- Search for a matching skill
- Create one if not found (after researching best practices)
- Ask where to store it (local or global)
- Execute the skill to complete your task
- Review and offer improvements based on execution
If you later find issues with the output and want to both fix the output AND improve the skill, invoke Skill Master again with your feedback:
Using command:
/skill-master please fix the business plan, the financial projections section needs more detail
Using natural language:
Invoke skill-master to adjust the market analysis - add competitor pricing data and export to PDF
Skill Master will:
- Fix your output immediately
- Link the feedback to the skill that created it
- Improve the skill so future executions are better
User Request
│
▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ SEARCH │────▶│ CREATE │────▶│ EXECUTE │
│ Skills │ │ if none │ │ Skill │
└─────────────┘ └─────────────┘ └─────────────┘
│
▼
┌─────────────┐
│ REVIEW │
│ (fresh agent)
└─────────────┘
│
┌───────────────────┴───────────────────┐
│ │
▼ ▼
No Divergence Divergence Found
│ │
▼ ▼
┌───────┐ ┌─────────────┐
│ DONE │ │ IMPROVE │
└───────┘ │ Skill │
└─────────────┘
The review phase is empirical, not theoretical:
- Compares what the skill says to do vs what actually happened
- Uses a fresh agent (via Task tool) for unbiased comparison
- Only suggests improvements when execution diverged from instructions
- User decides whether to apply improvements
This ensures skills improve based on real issues, not arbitrary criteria.
After a skill completes successfully, you might find issues with the output later:
- Content quality not meeting expectations
- Missing sections or details
- Format needs adjustment (e.g., export to PDF)
The skill executed perfectly (no divergences), but the output needs improvement.
Manually invoke Skill Master with your feedback to both fix the output AND improve the skill:
/skill-master please fix the market report, the competitive analysis section is too shallow
Or:
Invoke skill-master to adjust the documentation - add more code examples and export to PDF
User feedback request
│
▼
┌─────────────────────┐
│ FIX THE OUTPUT │ ◄── First, make the changes user requested
└─────────────────────┘
│
▼
┌─────────────────────┐
│ CHECK STATE FILE │ ◄── Look for .skill-master-state.json
└─────────────────────┘
│
├── State exists ──────────────────┐
│ ▼
│ ┌─────────────────────┐
│ │ REVIEW & IMPROVE │
│ │ - Link to skill │
│ │ - Capture feedback │
│ │ - Update skill │
│ └─────────────────────┘
│
└── No state ─────────────┐
▼
┌───────────┐
│ Done │
│ (fix only)│
└───────────┘
Skills are stateless - they don't have access to conversation history. When you say "fix the report", skill-master doesn't automatically know:
- Which skill created it
- What the original execution looked like
By explicitly invoking /skill-master, you signal that:
- This relates to a skill-generated output
- You want the skill improved, not just the output fixed
Skill Master maintains execution state in .skill-master-state.json:
{
"request": "create a market analysis report",
"skill_name": "market-research-reports",
"skill_path": ".claude/skills/market-research-reports",
"outputs": ["./reports/market-analysis.md"],
"state": "COMPLETE"
}This enables linking your feedback to the skill that produced the output.
Skills can be stored in two locations:
| Location | Scope | Use Case |
|---|---|---|
.claude/skills/ |
Project | Team-shared, committed to git |
~/.claude/skills/ |
Personal | Available across all projects |
Contributions welcome! Areas for improvement:
- Additional skill templates
- Better search algorithms
- More improvement patterns
- Integration with other agent frameworks
Skill Master is built for and relies on the following projects:
| Project | Description | Link |
|---|---|---|
| Claude Code | The AI coding agent platform this skill is built for | Anthropic |
| Anthropic Skills | Official skills repository (Tier 1 source) | anthropics/skills |
| K-Dense Scientific Skills | 139 scientific skills including market research (Tier 1 source) | K-Dense-AI/claude-scientific-skills |
| Awesome Claude Skills | Curated list of Claude skills (Tier 1 source) | ComposioHQ/awesome-claude-skills |
MIT License - See LICENSE