Two tools with clear verb-noun naming (get_, trigger_) and basic schemas. Both have descriptions and input validation via Zod. However, descriptions are minimal (34-56 chars, below the 50-200 char LLM-optimized baseline), parameter descriptions lack context about constraints and error recovery, and output schemas are not documented. No error guidance for LLMs (e.g., what to do if customer not found). Missing tool annotations (readOnlyHint, destructiveHint). Composition is sound but descriptions need expansion to meet production standards.

• Expand tool descriptions to 100-150 chars. Example: 'Retrieve a customer record by ID (format CUST-XXXXX). Returns name, email, and status. Use this to verify customer exists before processing refunds. If not found, try searching with a partial name.'
• Add descriptions to all parameters in the tool registration (not just in schemas). Example: 'customer_id: Customer ID in format CUST-XXXXX (5 digits). Must exist in the system or the call will fail.'
• Document output schemas in tool descriptions or as a separate schema field. Example: 'Returns {customerId, name, email, status} for get_customer_record; {refundId, customerId, amount, reason, status, processedAt} for trigger_refund.'
• Enhance error messages with recovery hints. When customer not found, return: 'CUSTOMER_NOT_FOUND: No customer with ID CUST-99999. Available customers: CUST-12345 (Alex Johnson), CUST-67890 (Priya Sharma). Try one of these or search with a partial name.'
• Add tool annotations in server.ts: register trigger_refund with destructiveHint=true and get_customer_record with readOnlyHint=true to guide LLM reasoning about idempotency and retry safety.