tool-design
Design Effective Tool APIs for AI Agents
こちらからも入手できます: ChakshuGautam,Asmayaseen,muratcankoylan
Poor tool design creates failure modes that no prompt engineering can fix. This skill provides principles for building tool interfaces optimized for AI agent reasoning and consumption.
スキルZIPをダウンロード
Claudeでアップロード
設定 → 機能 → スキル → スキルをアップロードへ移動
オンにして利用開始
テストする
「tool-design」を使用しています。 Review this tool: def search(query): "Search the database."
期待される結果:
- Issues found: (1) Vague name - search what? (2) Missing parameter docs - what format? (3) No return description (4) No error handling (5) No usage context
- Recommended: Rename to search_customers(query, fields), document query syntax, specify return format, add INVALID_QUERY and RATE_LIMITED error cases
「tool-design」を使用しています。 How do I reduce 50 specialized tools to a simpler set?
期待される結果:
- Apply consolidation: Group related tools by workflow, not by data entity
- Consider architectural reduction: Can file system access + standard utilities replace custom tools?
- Target 10-20 tools with namespacing: database.query, database.schema, web.search, web.fetch
セキュリティ監査
安全Static analysis flagged 73 potential issues but all are false positives. The SKILL.md file is documentation-only containing conceptual explanations and Python code examples in markdown blocks. No executable code, network calls, or system commands exist. External command patterns (8) are markdown code snippets, network patterns (3) are documentation URLs, cryptographic warnings (62) are text matches in prose, and reconnaissance flags (6) reference conceptual discussions.
品質スコア
作れるもの
Building Agent Tool Sets
Design comprehensive tool APIs for Claude or Codex to interact with your systems effectively
Debugging Tool Misuse
Diagnose why agents are failing to use tools correctly and improve descriptions
Reducing Tool Complexity
Apply architectural reduction to simplify over-engineered tool architectures
これらのプロンプトを試す
Review this tool specification and identify design issues: [paste tool code]. Check for vague descriptions, missing parameter documentation, unclear return formats, and error handling gaps.
This tool has unclear descriptions and causes agent failures: [paste tool]. Redesign it following tool design principles: clear what/when/returns, consolidated functionality, actionable error messages.
Analyze my current tool set: [list tools]. Identify opportunities for consolidation and reduction. Recommend a minimal set of general-purpose tools that could replace specialized ones.
I am building an agent for [domain]. Help me design a tool collection using consolidation and namespacing principles. Consider: distinct workflows, tool count limits (10-20), response format options, and recovery-oriented error messages.
ベストプラクティス
- Write tool descriptions that answer what it does, when to use it, and what it returns
- Prefer single comprehensive tools over multiple narrow tools to reduce ambiguity
- Design error messages that tell agents how to recover, not just what failed
回避
- Using vague descriptions like 'helps with' or 'can be used for'
- Naming tools with cryptic parameters like x, val, or data
- Building guardrail tools that constrain reasoning the model could handle