Five tools with basic verb-noun naming (get_*, execute_*) but significant quality gaps. All tools have descriptions (10-60 chars, below the 50-200 char baseline for LLM optimization). Input schemas present but minimal: execute_query and execute_query_json have only a 'query' string param with no format constraints, range limits, or validation guidance. get_table_columns and get_table_types accept only table_name with no description of expected format. Output schemas are completely undocumented, LLMs cannot infer what fields to expect from results. No error handling guidance (e.g., what happens on invalid SQL, missing table, connection failure?). No pagination support despite tools returning potentially large result sets. Tool descriptions lack WHEN to use each tool (e.g., when to call execute_query vs execute_query_json?) and any prerequisites. Per-tool average: 42/100.

• Add explicit output schema documentation to execute_query and execute_query_json. Example: 'Returns {columns: [{name, type, friendly_name}], rows: [[...]], row_count: int, total_count: int, has_more: bool}'.
• Enhance execute_query and execute_query_json descriptions to clarify use cases: 'Use execute_query for human-readable markdown tables (max 10 rows displayed). Use execute_query_json for structured data and tool chaining (returns full result set).'
• Add input constraints to query parameters: 'SQL query (max 10000 chars, SELECT/WITH only, no DDL/DML/TRUNCATE). Queries are read-only and executed against [database name].'
• Implement pagination: add limit (1-1000, default 50) and offset (default 0) parameters. Return total_count and has_more in output so LLMs can iterate.
• Add error classification to tool descriptions: 'On syntax error: returns 'Invalid SQL: [error detail]'. On timeout (>30s): retryable. On missing table: returns 'Table not found. Available tables: [list]'.