Server has 20 tools with consistent naming (verb_noun pattern) and descriptions present for all. However, descriptions are often generic and lack actionable context for LLM selection. Parameter schemas use Zod with type definitions, but descriptions are minimal (many under 50 chars). Output schemas are not documented, responses are JSON-stringified without field-level guidance. Error handling is absent; no recovery guidance or categorization. No tool annotations (readOnlyHint/destructiveHint). Portfolio tools (save/remove) lack confirmation patterns despite destructive risk.

• Document output schemas for all tools. For each tool, add a comment block describing returned fields, types, and structure. E.g., 'Returns: {funds: [{code: string, name: string, type: string, ...}], total: number}'.
• Expand parameter descriptions to 50-100 chars with actionable context. E.g., 'fundCode: 6-digit fund identifier (e.g., 161725). If you only have the fund name, call search_fund() first.' This guides LLM selection and prevents invalid calls.
• Add tool annotations to all tools. Use readOnlyHint: true for all get_* and search_* tools. Use destructiveHint: true for remove_portfolio. Use idempotentHint: true for save_portfolio (idempotent upsert).
• Implement error handling with recovery guidance. Wrap API calls in try-catch. Return structured errors: {error: string, code: string, suggestion: string}. E.g., 'Fund code not found. Try search_fund(keyword) to find the correct code.'
• Add confirmation pattern to remove_portfolio. Return {action: 'confirm_delete', fundCode: string, fundName: string} and require a second call with confirmed: true to actually delete. Prevents accidental data loss.