Three tools with complete schemas and Japanese descriptions, but descriptions are domain-specific and lack LLM-optimized guidance. All tools are READ_ONLY and well-scoped. Naming is clear (advise, compare_backends, list_traps). Parameters have types and defaults. However, descriptions lack actionable context for LLM selection (when to call each tool, what to do with results). Error handling in advise() returns structured errors but other tools lack recovery guidance. Output schemas are implicit rather than documented.

• Expand tool descriptions to 100-200 characters each, explaining WHEN to call the tool and WHAT the LLM should do with the result. Example: 'advise: Call this to answer specific RAG backend questions (e.g., "which backend has lowest p95 latency?"). Returns evidence chunks from rag-retriever-bench; synthesize your answer from these chunks only.'
• Document output schemas explicitly in tool descriptions or as a separate schema field. Example: 'Returns {question: str, evidence: [{source: str, text: str}], note: str}. Use evidence to ground your answer; do not cite sources outside this list.'
• Add error recovery guidance to compare_backends() and list_traps(). Example: 'If corpus_size is invalid, available sizes are returned; ask the user which size matches their use case.'
• Clarify parameter constraints in descriptions. For top_k, add 'Typical range: 3-10. Higher values increase latency but may improve answer quality.' For corpus_size, explain 'Only 10000 and 100000 are available; choose the size closest to your actual corpus.'
• Document tool composition in a README or tool group description. Explain: 'Use compare_backends() to see all measurements at a given corpus size; use advise() to answer specific questions about backend selection or tuning.'