Four tools with complete input schemas and descriptions, but quality varies significantly. search_ramen and vibe_search have detailed, LLM-optimized descriptions (150-200 chars) with well-structured parameters including enums and constraints. get_ramen_shop and get_ramen_changes are minimal (20-50 chars), lacking context for LLM selection. No output schemas documented anywhere in source. Parameters lack type definitions in schema (arrays/strings present but no 'type' field visible in JSON). Error handling and recovery guidance absent. No tool annotations (readOnlyHint, etc.) despite all being read-only operations.

• Document output schema for each tool. Example for search_ramen: {shops: [{shop_id, name, pref, keito[], spice_level, richness, hours, chain, venue_type, vibe_score}], fallback_shops?, fallback_reason?, total_count}. Include field types and descriptions.
• Expand get_ramen_shop and get_ramen_changes descriptions to 80-150 chars. Add WHEN guidance: 'Call after search_ramen to fetch full details (hours, menu, reviews) for a specific shop by ID' and 'Call to discover recent closures, reopenings, or menu changes in a prefecture.'
• Add enum constraints to schema. For 'pref': enum of all 47 romaji prefecture names. For 'keito': enum of 19 fine values (tonkotsu, iekei, jiro, miso, sapporo, shoyu, chuka_tanrei, kitakata, asahikawa, shio, tsukemen, tantanmen, abura_mazesoba, mazesoba, ramen_shop, curry, other, toripaitan, champon). For 'spice_level': enum [mild, medium, spicy]. For 'venue_type': enum [permanent, popup, all].
• Add error handling guidance to descriptions. Example: 'If no results found, returns fallback_shops with semantic matches. If shop_id invalid, returns error with available shop IDs in prefecture.' Include actionable recovery steps.
• Add readOnlyHint: true to all four tools in schema to signal to agents these are safe, non-destructive operations that can be retried without side effects.