Server has well-structured tool definitions with complete schemas and descriptions. All 5 tools have proper input schemas with typed parameters and enums. Descriptions are substantive (100-200 chars typical). However, output schemas are not documented, LLMs cannot predict response structure. Parameter descriptions are good but lack format/constraint details (e.g., 'limit' has no min/max bounds). No error handling guidance. Tools are read-only and well-composed, but lack actionable recovery patterns.

• Document output schemas for all tools. Example: 'Returns {results: [{id, title, description, category, relevance_score}], total_count, has_more}'. This enables LLMs to chain tools and extract fields correctly.
• Add min/max bounds to numeric parameters. Change 'limit' description to: 'Maximum number of results (1 - 50, default 10)'. Add bounds to any numeric param.
• Implement pagination for list operations. Add 'offset' and 'limit' params to get_design_principles, get_design_patterns, get_design_tokens. Return 'total_count' and 'has_more' in responses.
• Add error handling guidance in tool descriptions. E.g., 'If no results found, try broadening the search query or removing filters.' This guides LLM recovery.
• Specify ID format in parameter descriptions. E.g., 'id: kebab-case identifier (e.g., "visibility-of-system-status")'. Prevents hallucinated IDs.