Three tools with clear verb-noun naming (get_*, recommend_*) and documented descriptions. However, parameter descriptions are minimal (single short phrases), output schemas are not formally documented, and error handling lacks recovery guidance. Tools return JSON strings rather than structured objects, forcing LLMs to parse unstructured text. No input validation constraints (enums, ranges) visible. Descriptions average ~120 chars, within acceptable range but lack WHEN/WHY context. No tool annotations (readOnlyHint, etc.) despite all being read-only operations.

• Document output schemas for each tool. Define the structure of 'status', 'results', 'structured_matches', 'raw_text_excerpts', etc. with field types and descriptions. Return structured objects instead of JSON strings.
• Expand parameter descriptions to 50 - 150 chars. Include format hints: 'restaurant_name: partial or full name (e.g., "Iron" or "Sakura"); case-insensitive'.
• Add recovery guidance to error responses. E.g., 'No restaurant found. Try a partial name (e.g., "Iron") or use recommend_by_vibe to search by atmosphere.'
• Declare tool annotations via FastMCP. Mark all three tools with readOnlyHint=true so agents know they are safe to retry.
• Add pagination support to recommend_by_vibe. Cap structured_matches and raw_text_excerpts at 10 - 20 items; return a total_count and next_offset for large result sets.

Inferred effective spec: <=2025-11-25.

• Relies on Sampling (deprecated) - integrate directly with the LLM provider API