Three tools with clear, actionable descriptions (10 - 50 words each). All have input schemas with type definitions. However, output schemas are undocumented (no return type specs), parameter descriptions are minimal or absent, and error handling lacks recovery guidance. describe_schema and run_query have empty input objects or single params with no description text. monthly_close_summary accepts only one param with a format hint but no validation rules. No tool declares idempotency, permissions, or error categories. Naming is verb-forward (describe_, run_, monthly_) but generic. Overall: functional definitions suitable for read-only accounting queries, but missing LLM-optimization patterns (output schemas, param descriptions, error recovery) that would elevate to 70+.

• Add JSON Schema output definitions for all three tools. describe_schema returns a string of table names and columns, document as {type: 'string', description: 'Newline-separated list of tables and columns'}. run_query returns a formatted table or error, document as {type: 'string', description: 'Pipe-delimited table rows or error message'}. monthly_close_summary returns a formatted report, document as {type: 'string', description: 'Budget vs actual summary, sorted by variance'}.
• Add parameter descriptions and constraints. monthly_close_summary's 'month' param should read: 'Month in YYYY-MM format (e.g., 2026-05). Must be a valid month in the budget table.' run_query's 'sql' param should add: 'SELECT or WITH queries only. Max 500 rows returned; use LIMIT or aggregation to narrow results.'
• Enhance error messages with recovery guidance. Instead of 'REFUSED: only SELECT/WITH queries are allowed', return 'Invalid query: only SELECT/WITH are allowed. Try: SELECT * FROM accounts LIMIT 10'. Instead of 'SQL ERROR: ...', return 'Query failed: [error]. Check table names with describe_schema() first.'
• Declare tool annotations in the schema. Add readOnlyHint: true and idempotentHint: true to all three tools so LLMs know they are safe to retry and do not modify state.
• Document the 500-row and 40KB caps in run_query's description and output schema. Add a note: 'Results are capped at 500 rows and 40,000 characters. Use aggregation, filtering, or LIMIT to narrow results if truncated.'