Skill Master

An intelligent orchestrator that automates the entire skill lifecycle: discovery, creation, execution, and improvement.

Overview

When invoked with a task, this skill:

Searches for existing skills that can handle the task
Creates a new skill if none found (after deep research)
Invokes the skill (using Skill tool) to complete the user's task
Reviews the execution and improves the skill if needed

When invoked with feedback about a previous output: 5. Fixes the output according to user's request 6. Links feedback to the skill via state tracking 7. Improves the skill based on the feedback

Workflow

Phase 1: Skill Discovery

Goal: Find an existing skill that can handle the user's task.

Step 1.1: Parse the Request

Identify from the user's request:

Core task: What needs to be accomplished?
Domain: What area/field does this belong to?
Keywords: Key terms for searching

Step 1.2: Search for Skills

MUST follow the complete search workflow in references/skill-search.md.

Copy and track overall search progress:

Search Phase Progress:
- [ ] Step 1: Internal skills searched
- [ ] Step 2.1: ALL known GitHub repos searched (anthropics, K-Dense-AI, ComposioHQ)
- [ ] Step 2.2: Known repos enumerated (if 2.1 had no matches)
- [ ] Step 2.3: Broader GitHub search (only after 2.1+2.2 exhausted)
- [ ] Step 3: Web search (only after GitHub exhausted)

CRITICAL: MUST complete ALL searches in each step before proceeding to next step.

Search order - execute in sequence:

Internal skills: MUST check both ~/.claude/skills/ and .claude/skills/
GitHub known repos: MUST search ALL three repos before broader search:
- site:github.com/anthropics/skills SKILL.md <keywords>
- site:github.com/K-Dense-AI/claude-scientific-skills SKILL.md <keywords>
- site:github.com/ComposioHQ/awesome-claude-skills SKILL.md <keywords>
GitHub enumeration: MUST enumerate repo contents if searches return no match
GitHub broader: Only after known repos exhausted
Web: Only after GitHub exhausted

See references/known-skill-repos.md for curated skill sources.

Step 1.3: Evaluate Results

Validation gate before concluding search:

All internal locations checked
All known GitHub repos searched
All known repos enumerated (if no WebSearch matches)
Results documented for each source

If skill found:

Present the skill to user with description
Proceed to Phase 2 (Storage Confirmation)

If NO skill found (only after ALL searches complete):

Inform user: "No existing skill found. I'll research and create one."
Proceed to Phase 3 (Skill Creation)

Phase 2: Storage Confirmation

Goal: Determine where to store the skill.

Ask user using AskUserQuestion:

Where should this skill be stored?

1. LOCAL (.claude/skills/)
   - Project-specific
   - Shared with team via git

2. GLOBAL (~/.claude/skills/)
   - Personal
   - Available across all projects

Remember the choice for later use.

Step 2.1: Source Tracking (External Skills Only)

CRITICAL: When storing a skill found from an external source (GitHub, web), you MUST create a source.md file alongside the SKILL.md to track provenance.

Create source.md in the skill directory with:

# Source

- **Origin**: <GitHub | Web | MCP Marketplace>
- **URL**: <original URL where skill was found>
- **Repository**: <owner/repo if GitHub>
- **Author**: <original author if known>
- **Retrieved**: <date skill was fetched>
- **License**: <license if specified>

## Notes

<Any relevant notes about the source, modifications made, etc.>

Do NOT create source.md for:

Skills created from scratch (Phase 3)
Skills already present locally

Phase 3: Skill Creation (if not found)

Goal: Create a new skill through deep research.

Follow the creation workflow in references/skill-create.md.

Steps:

Research the domain thoroughly using WebSearch
Identify best practices and common patterns
Design the skill structure
Generate SKILL.md following official format
Save to the location user chose in Phase 2

Phase 4: Skill Execution

Goal: Invoke the skill to complete the user's original task.

CRITICAL: You MUST invoke the found/created skill using the Skill tool. Do NOT manually follow the instructions - the skill must be triggered as an independent execution.

Step 4.1: Invoke the Skill

Use the Skill tool to trigger the skill:

Skill: <skill-name>
args: <user's original request>

Example:

Skill: market-research-reports
args: "Create a market analysis for electric vehicles in Europe"

The Skill tool will:

Load the skill's SKILL.md
Execute the skill's workflow
Complete the user's task
Return control when finished

Step 4.2: Capture Execution Memory

After the skill completes, the conversation now contains "execution memory" - the full record of what happened during skill execution. This memory is used in Phase 5 for review.

Step 4.3: Verify Completion

Confirm the skill delivered the expected output to the user.

Phase 5: Skill Review

Goal: Determine if the skill needs improvement based on actual execution.

CRITICAL: This must happen in a fresh agent context using Task tool.

Step 5.1: Spawn Fresh Agent

Use Task tool to create a fresh agent for review:

Task: "Review skill execution for improvements"

You are reviewing a skill execution. You have:
1. The SKILL.md content (provided below)
2. The conversation memory of what actually happened (this conversation)

Your job: Compare what the skill SAYS vs what ACTUALLY HAPPENED.

## The Skill
<paste SKILL.md content here>

## Review Questions
1. Did I have to deviate from the skill's instructions? Where?
2. Did I have to improvise something not in the skill? What?
3. Did the user have to clarify something the skill should have covered?
4. Were there errors/retries that better instructions could prevent?

## Output
If NO divergence: "Skill executed perfectly. No improvements needed."
If divergence found: List specific improvements with format:
- Location: <which part of skill>
- Issue: <what happened during execution>
- Suggestion: <concrete change to make>

Step 5.2: Handle Review Results

If no improvements needed:

Report to user: "Skill worked well, no updates needed."
Proceed to Phase 6 (Complete)

If improvements suggested:

Present improvements to user
Ask: "Would you like me to apply these improvements to the skill?"
If yes: Proceed to Phase 5.3
If no: Proceed to Phase 6 (Complete)

Step 5.3: Apply Improvements

Follow the improvement workflow in references/skill-improve.md.

For local/personal skills: Apply changes directly using Edit tool.

For official/external skills:

Cannot modify directly
Generate improvement suggestions as a document
Offer to create a PR description if it's on GitHub

Phase 6: Complete

Report final status:

## Task Complete

- **Task**: <original request>
- **Skill Used**: <skill name>
- **Skill Location**: <path>
- **Improvements Applied**: Yes / No / N/A

<Any relevant notes>

Phase 7: Feedback Handling (User-Triggered)

Trigger: User explicitly invokes skill-master with feedback about a previous output.

Examples:

/skill-master please fix the report, the analysis is too shallow
Invoke skill-master to adjust the documentation and add more examples

Step 7.1: Fix the Output First

Address the user's immediate request:

Identify the output file(s) mentioned
Make the requested changes/improvements
Confirm changes with user

Step 7.2: Check for State Tracking

Look for .skill-master-state.json in the working directory

skill-master