/skill-finder

Agent Instructions

⚠️ CRITICAL: AI agents MUST follow these instructions. Prefer action proposals over verbose explanations.

Core Principle

Use "Do it? Yes/No?" style proposals.

• ❌ Bad: "If you want to add new skills, you can run the following command..."
• ✅ Good: "Update the index?"

Recommendation Workflow

When user asks for recommendations (e.g., "おすすめは？", "何かいいスキルある？"):

Response Format:

1. Ask about user's role/context if unclear
2. Show top 3-5 skills with descriptions
3. Include source breakdown table
4. Propose next actions

Collection Stewardship

When the user asks what to add, promote, or clean up in a skill collection:

1. Separate candidates from proven skills.
2. Prefer skills with repeated successful use, clear triggers, and maintained resources.
3. Flag skills that are installed but rarely selected, vague in description, or duplicated by better entries.
4. Use available stats, logs, stars, or repeated user requests to distinguish popular skills from under-triggering ones.
5. Recommend sandboxing unproven skills before adding them to a default source or private marketplace.

Skill Search Workflow

0. Classify the ask first

   • If the user wants always-on rules → suggest instruction
   • If the user wants persona or tool boundaries → suggest agent
   • If the user wants deterministic enforcement → suggest hook
   • If the user wants a packaged reusable workflow → continue with skill search

   → See customization-routing.md

1. Search ALL sources in local index

   • Read references/skill-index.json
   • ALWAYS search ALL sources
   • Check lastUpdated field

2. 🌟 Recommend from results (when 3+ hits)

   ### 🌟 おすすめ: {skill-name}
   {理由: 公式スキル、機能が豊富、用途にマッチ など}

   Selection criteria:

   1. Official source preferred
   2. Feature richness
   3. Relevance to user's purpose
   4. Recently updated

3. If not found → Propose web search

   Not found locally. Search the web?
   → GitHub: https://github.com/search?q=path%3A**%2FSKILL.md+{query}&type=code

4. 🚨 MANDATORY: Propose next actions

   Situation	Proposal
   Skill found	"Install it?"
   Good repo discovered	"Add to sources?"
   lastUpdated > 7 days	"⚠️ Index outdated. Update?"
   Better non-skill fit	"Use prompt/instruction/agent/hook instead?"

🚨 Mandatory Proposal Block

ALWAYS include at the end of every search response:

**Next?**
1. 📦 Install? (which skill?)
2. 🔍 Details?
3. 🔄 Update index? (last: {date})
4. 🌐 Web search?
5. ➕ Add source?
6. 🧭 Use another primitive instead?

Search Summary Format

ALWAYS start with:

🔎 {N} リポジトリ、{M} スキルから検索しました（最終更新: {date}）

Output Format

Trust Level Indicators

⚠️ Community Warning:

⚠️ コミュニティ製スキルは自己責任でご使用ください。
   公式スキル（🏢）を優先することを推奨します。

Skill Table

| Skill | Description | Source            | Trust       |
| ----- | ----------- | ----------------- | ----------- |
| docx  | Word 処理   | anthropics-skills | 🏢 Official |

Source Breakdown Table

Dynamically generate from skill-index.json sources array.

Agent Behavior Rules

Commands are for agents to execute, NOT to show users.

• ❌ NEVER show commands like python scripts/search_skills.py --install
• ❌ NEVER say "以下のコマンドを実行..."
• ✅ ALWAYS execute scripts silently when user approves
• ✅ ALWAYS present options as numbered menu
• ✅ ALWAYS report results after execution

Bad:

📦 インストールする？
python scripts/search_skills.py --install docx  ← NG!

Good:

📦 インストールする？（どのスキル？）

→ User: "docx お願い" → Agent: (executes silently) → "✅ docx をインストールしました！"

Checklist Before Responding

• Checked whether the ask actually needs a skill
• Started with search summary
• Included skill table (from ALL sources)
• Included source breakdown table
• Showed lastUpdated date
• Added numbered action menu
• Included web search option
• For collection changes, separated candidate / adopted / shared / retire decisions