Critical gaps across naming, descriptions, and schemas. 30 tools total: 8 hand-crafted (advertiser, adgroups, creatives, audiences, funds) with partial schemas; 22 auto-generated with minimal, generic descriptions and vague 'params' object schemas. Hand-crafted tools have better structure but still lack comprehensive parameter documentation. Auto-generated tools (tools 24 - 30) use a catch-all 'params: object' pattern that defeats schema validation and LLM guidance. Descriptions are often single-phrase Chinese translations without actionable context. No error handling guidance, no output schemas documented, no tool annotations.

• Replace auto-generated 'params: object' with explicit field schemas. Parse api-endpoints.json to extract actual parameter names, types, and constraints. Generate proper Zod schemas with .describe() for each field.
• Expand all descriptions to 50 - 200 characters with WHAT, WHEN, and WHAT-IT-RETURNS. Example: 'Get advertiser account info (name, budget, status). Call this first to verify account access before creating ad groups. Returns account_id, corporation_name, daily_budget.'
• Document output schemas in tool descriptions or as separate schema comments. Specify which fields are returned and their types so LLMs can chain tools. Example: 'Returns: {account_id: number, corporation_name: string, daily_budget: number}'.
• Add tool annotations: mark destructive tools with destructiveHint=true, read-only tools with readOnlyHint=true, and idempotent tools with idempotentHint=true. This guides LLM retry logic.
• For enum parameters (marketing_goal, configured_status, bid_mode), replace free-form descriptions with explicit Zod enums: z.enum(['MARKETING_GOAL_PRODUCT_SALES', 'MARKETING_GOAL_LEAD_GENERATION', ...]).