Skill Finder - Search Skills

Setup

If ~/skill-finder/ doesn't exist or is empty, tell the user you are initializing local memory files, then follow setup.md.

When to Use

User asks to find a skill, discover capabilities, or wonders if something exists. Handles searching, evaluating quality, comparing options, and learning what the user values.

Architecture

Memory lives in ~/skill-finder/. See memory-template.md for structure.

~/skill-finder/
├── memory.md     # Preferences + liked/passed skills
└── searches.md   # Recent search history (optional)

Quick Reference

Core Rules

1. Search by Need, Not Name

User says "help with PDFs" — think about what they actually need:

• Edit? → clawhub search "pdf edit"
• Create? → clawhub search "pdf generate"
• Extract? → clawhub search "pdf parse"

2. Evaluate Before Recommending

Never recommend blindly. Check evaluate.md criteria:

• Description clarity
• Download count (popularity = maintenance)
• Last update (recent = active)
• Author reputation

3. Present with Reasoning

Don't just list skills. Explain why each fits:

"Found pdf-editor — handles form filling and annotations, 2.3k downloads, updated last week. Matches your need for editing contracts."

4. Learn Preferences

When user explicitly states what they value, confirm and update ~/skill-finder/memory.md:

• "I prefer minimal skills" → add to Preferences
• "This one is great" → add to Liked with reason
• "Too verbose" → add to Passed with reason

Do not infer hidden preferences from behavior-only signals.

5. Check Memory First

Before recommending, read memory.md:

• Skip skills similar to Passed ones
• Favor qualities from Liked ones
• Apply stated Preferences

6. Never Bypass Security Warnings

If a candidate skill is marked risky by scanner output:

• Explain the warning and inspect details first
• Prefer a safer alternative
• Do not run force-install flags for the user
• Install only normal, non-risky options with explicit user consent

Search Commands

# Primary search
npx clawhub search "query"

# Install (with user consent)
clawhub install <slug>

# Get skill details
clawhub inspect <slug>

# See what's installed
clawhub list

Workflow

1. Understand — What does user actually need?
2. Search — Try specific terms first, broaden if needed
3. Evaluate — Check quality signals (see evaluate.md)
4. Compare — If multiple match, rank by fit + quality
5. Recommend — Top 1-3 with clear reasoning
6. Learn — Store explicit feedback in memory

Common Traps