Server has 4 tools with basic verb-noun naming (get_*). All tools have descriptions in Japanese, but they are extremely brief (10-20 chars) and lack actionable context for LLM selection. Input schemas are minimal: 3 tools accept a single 'name' string parameter with no type validation, enum constraints, or format guidance; 2 tools have empty input objects. No output schemas are documented, LLMs cannot predict response structure. No error handling guidance, no pagination support despite list tools, no parameter validation rules. Tools return unstructured string concatenation rather than structured JSON. No security considerations, no idempotency markers, no permission gates.

• Expand tool descriptions to 50-200 characters. Example for get_resonator: 'Retrieve detailed information about a specific resonator by name, including attributes, weapon type, nation, and story content. Use this when you need full details about one resonator; use get_resonators() to browse all available resonators.'
• Document output schemas as JSON objects with typed fields. Example: get_resonator returns {name: string, attribute: string, weaponType: string, nation: string, stories: [{title: string, content: string}]}. Return structured JSON, not string concatenation.
• Add parameter constraints to 'name' inputs: description should state 'The exact name of the resonator (case-sensitive, e.g., "Jiyan", "Encore"). Call get_resonators() first to see available names if unsure.'
• Add pagination to list tools: accept optional 'limit' (1-100, default 20) and 'offset' (default 0) parameters. Return {items: [...], total: number, hasMore: boolean} so LLMs can fetch additional pages.
• Wrap GraphQL calls in try-except blocks and return structured error responses: {error: string, suggestion: string}. Example: 'Resonator "InvalidName" not found. Available resonators: Jiyan, Encore, Aalto. Try one of these or call get_resonators() to see all options.'