メニュー
Use when building "MCP server", "Model Context Protocol", creating "Claude tools", "MCP tools", or asking about "FastMCP", "MCP SDK", "tool development for LLMs", "external API integration for Claude"
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when evaluating agent performance, building test frameworks, measuring quality, or asking about "agent evaluation", "LLM-as-judge", "agent testing", "quality metrics", "evaluation rubrics", "agent benchmarks"
Use when "codex", "use gpt", "gpt-5", "openai codex", "let openai", "full-auto", "autonomous code generation"
Use when compressing agent context, implementing conversation summarization, reducing token usage in long sessions, or asking about "context compression", "conversation history", "token optimization", "context limits", "summarization strategies"
Use when diagnosing agent failures, debugging lost-in-middle issues, understanding context poisoning, or asking about "context degradation", "lost in middle", "context poisoning", "attention patterns", "context clash", "agent performance drops"
Use when optimizing agent context, reducing token costs, implementing KV-cache optimization, or asking about "context optimization", "token reduction", "context limits", "observation masking", "context budgeting", "context partitioning"
Use when "CrewAI", "multi-agent systems", "agent orchestration", "AI crews", or asking about "autonomous agents", "agent collaboration", "role-based agents", "agent workflows", "AI team coordination"
| name | mcp-development |
| description | Use when building "MCP server", "Model Context Protocol", creating "Claude tools", "MCP tools", or asking about "FastMCP", "MCP SDK", "tool development for LLMs", "external API integration for Claude" |
| version | 1.0.0 |
Build high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services.
| Principle | Why |
|---|---|
| Consolidate operations | Single tool for complete tasks |
| Return high-signal data | Agents have limited context |
| Provide format options | "concise" vs "detailed" modes |
| Use human-readable IDs | Not technical codes |
| Make errors actionable | Guide toward correct usage |
Key concept: Don't just wrap API endpoints. Design tools that enable complete workflows agents actually need.
| Step | Action |
|---|---|
| Study MCP Protocol | Read modelcontextprotocol.io/llms-full.txt |
| Study SDK docs | Python or TypeScript SDK README |
| Study target API | Read ALL available documentation |
| Create implementation plan | Before writing code |
| Decision | Options |
|---|---|
| Language | Python (FastMCP) or TypeScript |
| Tool granularity | Atomic vs workflow-oriented |
| Response format | JSON, Markdown, or both |
| Error handling | What errors can occur, how to recover |
| Component | Purpose |
|---|---|
| Input validation | Pydantic (Python) or Zod (TypeScript) |
| Tool descriptions | Clear, with examples |
| Error messages | Include suggested next steps |
| Response formatting | Consistent across tools |
Critical: MCP servers are long-running processes. Never run directly in main process.
| Approach | How |
|---|---|
| Evaluation harness | Recommended |
| tmux session | Run server separately |
| Timeout wrapper | timeout 5s python server.py |
| MCP Inspector | Official debugging tool |
| Annotation | Meaning | Default |
|---|---|---|
| readOnlyHint | Doesn't modify state | false |
| destructiveHint | Can cause damage | true |
| idempotentHint | Repeated calls safe | false |
| openWorldHint | Interacts externally | true |
Key concept: Annotations help the LLM decide when and how safely to use tools.
| Pattern | Use Case |
|---|---|
| Required fields | Core parameters |
| Optional with defaults | Convenience parameters |
| Enums | Limited valid values |
| Min/max constraints | Numeric bounds |
| Pattern matching | Format validation (email, URL) |
| Good | Bad | Why |
|---|---|---|
user_email | e | Self-documenting |
limit | max_results_to_return | Concise but clear |
include_archived | ia | Descriptive boolean |
| Format | Use Case |
|---|---|
| JSON | Programmatic use, structured data |
| Markdown | Human readability, reports |
| Hybrid | JSON in markdown code blocks |
| Guideline | Why |
|---|---|
| ~25,000 token limit | Context constraints |
| Truncate with indicator | Don't silently cut |
| Support pagination | limit and offset params |
| Include metadata | Total count, has_more |
| Element | Purpose |
|---|---|
| What failed | Clear description |
| Why it failed | Root cause if known |
| How to fix | Suggested next action |
| Example | Correct usage |
Key concept: Error messages should guide the agent toward correct usage, not just diagnose problems.
| Check | Description |
|---|---|
| No duplicated code | Extract shared logic |
| Consistent formats | Similar ops return similar structure |
| Full error handling | All external calls wrapped |
| Type coverage | All inputs/outputs typed |
| Comprehensive docstrings | Every tool documented |
| Check | Description |
|---|---|
| Clear descriptions | Model knows when to use |
| Good examples | In docstring |
| Sensible defaults | Reduce required params |
| Consistent naming | Group related with prefixes |
| Practice | Why |
|---|---|
| One tool = one purpose | Clear mental model |
| Comprehensive descriptions | LLM selection accuracy |
| Include examples in docstrings | Show expected usage |
| Return actionable errors | Enable self-correction |
| Test with actual LLM | Real-world validation |
| Version your server | Track compatibility |
