skill-finder（技能查找器）

When to Use

User asks how to do something, wonders whether a skill exists, wants a new capability, or asks for the best skill for a job. Use before solving manually when an installable skill could extend the agent, replace a weak skill, or offer a safer alternative.

Architecture

Memory lives in ~/skill-finder/. If ~/skill-finder/ does not exist or is empty, run setup.md.

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

Migration

If upgrading from a previous version, see migration.md for data migration steps. The agent MUST check for legacy memory structure before proceeding.

Quick Reference

Activation Signals

Activate when the user says things like:

• "How do I do X?"
• "Is there a skill for this?"
• "Can you do this better?"
• "Find a skill for X"
• "I need a safer or more maintained option"
• "What should I install for this task?"

Also activate when the user describes a missing capability, a repetitive workflow, or frustration with a current skill.

Search Sources

This skill can search two ecosystems:

Default mode: search both sources, then compare results together.

Configurable modes:

• both — recommended default
• clawhub — only search ClawHub
• skills.sh — only search the Skills.sh ecosystem

Store the current mode in ~/skill-finder/memory.md. If the user has no saved preference yet, explain the two sources once, recommend both, and save the explicit choice.

Security Note

This skill uses npx clawhub and npx skills to discover and install skills from two different ecosystems. Review candidates before installation, keep installs opt-in, and keep the source attached to every recommendation.

Data Storage

This skill stores local preference data in ~/skill-finder/:

• Source mode, explicit preferences, liked skills, and passed skills in the local memory file inside ~/skill-finder/
• Optional recent search history in a local search log inside ~/skill-finder/

Create on first use: mkdir -p ~/skill-finder

Core Rules

1. Search Both Sources by Default

Unless the user has explicitly chosen otherwise, search ClawHub and Skills.sh for the same need, then compare the strongest results together.

Never assume a Skills.sh result can be installed with clawhub, or the reverse. Keep the source and install command attached to every recommendation.

2. Trigger on Capability Gaps, Not Just Explicit Search Requests

Do not wait only for "find a skill." Activate when the user describes missing functionality, asks how to do a task faster, or wants a better tool for a job.

3. Search by Need, Not Name

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

• Edit? -> npx clawhub search "pdf edit" and npx skills find pdf edit
• Create? -> npx clawhub search "pdf generate" and npx skills find pdf generate
• Extract? -> npx clawhub search "pdf parse" and npx skills find pdf parse

4. Evaluate Before Recommending

Never recommend blindly. Inspect strong candidates and check evaluate.md criteria:

• Description clarity
• Download count (popularity = maintenance)
• Last update (recent = active)
• Author or repository reputation
• Install scope and friction

For Skills.sh candidates, pay attention to the package source and install string the CLI returns.

5. Present a Decision, Not a Dump

Don't just list skills. Explain why each fits, who it is best for, and why the winner wins:

"Best fit: pdf-editor from ClawHub — handles form filling and annotations, 2.3k downloads, updated last week. Matches your need for editing contracts better than the Skills.sh options."

When there are multiple good fits, rank the top 1-3 and call out tradeoffs clearly.

6. Learn Preferences and Source Mode

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

• "Search both by default" -> set source mode to both
• "Only use Skills.sh for this workspace" -> set source mode to skills.sh
• "Only check ClawHub" -> set source mode to clawhub
• "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.

7. Check Memory First

Before recommending, read memory.md:

• Respect saved source mode unless the user overrides it
• Skip skills similar to Passed ones
• Favor qualities from Liked ones
• Apply stated Preferences

8. Respect Installation and Security Boundaries

If a candidate skill is marked risky by scanner output, or the install path is unclear:

• Explain the warning or ambiguity first
• Prefer a safer alternative
• Do not run force-install flags for the user
• Do not auto-accept install prompts with -y
• Do not choose global install scope unless the user explicitly wants it
• Install only with explicit user consent

9. Fallback Gracefully

If nothing is strong enough:

• Say what was searched
• Say which source mode was used
• Explain why the matches are weak
• Help directly or suggest creating a purpose-built skill

Search Commands

# ClawHub search and inspect
npx clawhub search "query"
npx clawhub inspect <slug>
npx clawhub install <slug>
npx clawhub list

# Skills.sh ecosystem
npx skills find [query]
npx skills add <owner/repo@skill>
npx skills list
npx skills check
npx skills update

# Example install string returned by `npx skills find`
npx skills add vercel-labs/agent-skills@vercel-react-best-practices

Workflow

1. Detect - Is the user describing a capability gap or installable need?
2. Load memory - Read ~/skill-finder/memory.md for source mode and preferences
3. Understand - What does user actually need?
4. Search - Use both by default, or the saved single-source mode
5. Evaluate - Check quality signals (see evaluate.md)
6. Compare - Rank results across both sources by fit + quality
7. Recommend - Top 1-3 with clear reasoning and a winner
8. Install or fallback - Install only with consent, otherwise help directly
9. Learn - Store explicit feedback in memory

Recommendation Format

When presenting results, prefer this structure:

Best fit: <slug or owner/repo@skill>
Source: <ClawHub or Skills.sh>
Why it wins: <1-2 lines>
Install: <exact command>
Tradeoffs: <what it does not cover or where alternative is stronger>
Alternatives: <slug>, <slug>
Next step: Install now or continue without installing

Common Traps

• Waiting for the exact phrase "find a skill" -> misses proactive discovery moments
• Searching generic terms -> gets noise. Be specific: "react testing" not "testing"
• Searching only one ecosystem when the saved mode is both
• Recommending by name match only -> misses better alternatives with different names
• Mixing install commands between ClawHub and Skills.sh
• Ignoring download counts -> low downloads often means abandoned
• Not checking last update -> outdated skills cause problems

Security & Privacy

Data that leaves your machine:

• Search queries sent to ClawHub registry (public search)
• Search queries sent through the skills CLI / Skills.sh ecosystem

Data that stays local:

• All preferences in ~/skill-finder/memory.md
• Search history (if enabled)

This skill does NOT:

• Install skills without user consent
• Use force-install flags to skip scanner warnings
• Auto-confirm npx skills add with -y
• Switch to global install scope silently
• Collect hidden behavior data
• Access files outside ~/skill-finder/

Related Skills

Install with npx clawhub install <slug> if user confirms:

• skill-manager — manages installed skills, suggests updates
• skill-builder — creates new skills from scratch
• skill-update — updates existing skills

Feedback

• If useful: clawhub star skill-finder
• Stay updated: clawhub sync