Three tools with basic verb-noun naming (get_*, apply_*) and present descriptions, but significant gaps in parameter documentation and output schema clarity. All tools lack parameter descriptions in the schema, only tool names and employee_id/leave_dates are typed, but no descriptions explain what each parameter controls or constrains. Output is unstructured plain text (e.g., 'Employee ID not found.'), not structured JSON. No error handling guidance for LLMs. apply_leave is a write operation but lacks idempotent/destructive hints. Descriptions are 40 - 80 chars, below the 50 - 200 char ideal for LLM optimization.

• Add parameter descriptions to all tools. Example: employee_id: 'The unique employee identifier (e.g., E001, E002). Must be a valid employee in the system.' leave_dates: 'Array of ISO 8601 date strings (YYYY-MM-DD) for which leave is requested. Dates must be in the future and not already taken.'
• Return structured JSON from all tools instead of plain text. Example: get_leave_balance should return {"employee_id": "E001", "balance": 18, "unit": "days"} instead of 'E001 has 18 leave days remaining.'
• Add tool annotations to apply_leave: mark it with destructiveHint=true and idempotentHint=false so LLMs know it modifies state and cannot be safely retried without checking side effects.
• Enhance error messages with recovery guidance. Example: 'Employee ID "E999" not found. Valid employees: E001, E002. Try searching with a partial name or contact HR.' This enables LLM self-correction.
• Document in tool descriptions that the database is in-memory and ephemeral (for demo/testing only). Add a note: 'This tool uses a mock database that resets on server restart. For production, integrate with a persistent leave management system.'