Extract

/self-improving-agent:extract — Create Skills from Patterns

Transforms a recurring pattern or debugging solution into a standalone, portable skill that can be installed in any project.

Usage

/self-improving-agent:extract <pattern description>                  # Interactive extraction
/self-improving-agent:extract <pattern> --name docker-m1-fixes       # Specify skill name
/self-improving-agent:extract <pattern> --output ./skills/            # Custom output directory
/self-improving-agent:extract <pattern> --dry-run                     # Preview without creating files

When to Extract

A learning qualifies for skill extraction when ANY of these are true:

Workflow

Step 1: Identify the pattern

Read the user's description. Search auto-memory for related entries:

MEMORY_DIR="$HOME/.claude/projects/$(pwd | sed 's|/|%2F|g; s|%2F|/|; s|^/||')/memory"
grep -rni "<keywords>" "$MEMORY_DIR/"

If found in auto-memory, use those entries as source material. If not, use the user's description directly.

Step 2: Determine skill scope

Ask (max 2 questions):

• "What problem does this solve?" (if not clear)
• "Should this include code examples?" (if applicable)

Step 3: Generate skill name

Rules for naming:

• Lowercase, hyphens between words
• Descriptive but concise (2-4 words)
• Examples: docker-m1-fixes, api-timeout-patterns, pnpm-workspace-setup

Step 4: Create the skill files

Spawn the skill-extractor agent for the actual file generation.

The agent creates:

<skill-name>/
├── SKILL.md            # Main skill file with frontmatter
├── README.md           # Human-readable overview
└── reference/          # (optional) Supporting documentation
    └── examples.md     # Concrete examples and edge cases

Step 5: SKILL.md structure

The generated SKILL.md must follow this format:

---
name: <skill-name>
description: "<one-line description>. Use when: <trigger conditions>."
---

# <Skill Title>

> One-line summary of what this skill solves.

## Quick Reference

| Problem | Solution |
|---------|----------|
| {{problem 1}} | {{solution 1}} |
| {{problem 2}} | {{solution 2}} |

## The Problem

{{2-3 sentences explaining what goes wrong and why it's non-obvious.}}

## Solutions

### Option 1: {{Name}} (Recommended)

{{Step-by-step with code examples.}}

### Option 2: {{Alternative}}

{{For when Option 1 doesn't apply.}}

## Trade-offs

| Approach | Pros | Cons |
|----------|------|------|
| Option 1 | {{pros}} | {{cons}} |
| Option 2 | {{pros}} | {{cons}} |

## Edge Cases

- {{edge case 1 and how to handle it}}
- {{edge case 2 and how to handle it}}

Step 6: Quality gates

Before finalizing, verify:

• SKILL.md has valid YAML frontmatter with name and description
• name matches the folder name (lowercase, hyphens)
• Description includes "Use when:" trigger conditions
• Solutions are self-contained (no external context needed)
• Code examples are complete and copy-pasteable
• No project-specific hardcoded values (paths, URLs, credentials)
• No unnecessary dependencies

Step 7: Report

✅ Skill extracted: {{skill-name}}

Files created:
  {{path}}/SKILL.md          ({{lines}} lines)
  {{path}}/README.md         ({{lines}} lines)
  {{path}}/reference/examples.md  ({{lines}} lines)

Install: /plugin install (copy to your skills directory)
Publish: clawhub publish {{path}}

Source: MEMORY.md entries at lines {{n, m, ...}} (retained — the skill is portable, the memory is project-specific)

Examples

Extracting a debugging pattern

/self-improving-agent:extract "Fix for Docker builds failing on Apple Silicon with platform mismatch"

Creates docker-m1-fixes/SKILL.md with:

• The platform mismatch error message
• Three solutions (build flag, Dockerfile, docker-compose)
• Trade-offs table
• Performance note about Rosetta 2 emulation

Extracting a workflow pattern

/self-improving-agent:extract "Always regenerate TypeScript API client after modifying OpenAPI spec"

Creates api-client-regen/SKILL.md with:

• Why manual regen is needed
• The exact command sequence
• CI integration snippet
• Common failure modes

Tips

• Extract patterns that would save time in a different project
• Keep skills focused — one problem per skill
• Include the error messages people would search for
• Test the skill by reading it without the original context — does it make sense?