Server defines 5 tools with consistent verb_noun naming (get_*) and Chinese descriptions. All tools have input schemas with Zod validation and parameter descriptions. However, descriptions are brief (10-20 chars in English equivalent) and lack actionable context for LLM selection. No output schemas documented. Parameters are all optional strings with minimal constraint guidance. Error handling is basic (isError flag only). No tool annotations (readOnlyHint, destructiveHint). Composition is sound, each tool has one responsibility and tools chain via natural identifiers (county, city, river_name). Missing: pagination guidance, result limits, recovery hints in error responses.

• Document output schema for each tool. Example for get_flood_potential: {flood_zones: [{area_name, risk_level: 'low'|'medium'|'high', affected_population}], timestamp: ISO8601}
• Expand descriptions to 50-150 chars with WHEN/WHY context. E.g., 'Retrieve real-time flood alerts from Civil IoT sensors. Use when monitoring active weather events. Returns alert location, severity, and timestamp.'
• Add enum constraints for county/city parameters. Define valid values in schema or description: 'county: one of [臺北市, 新北市, 桃園市, ...]'
• Implement pagination: add optional limit (1-100, default 20) and offset/cursor parameters to get_flood_warnings and get_rainfall_data. Document in description: 'Returns up to 20 results; use offset for pagination.'
• Enhance error messages with recovery hints. When validation fails, return: {isError: true, content: 'Invalid county: got "台北市". Did you mean: 臺北市? Valid counties: [list]'}