Server has 21 tools with complete Zod schemas and descriptions in Korean. Naming follows verb_noun convention (list_, create_, set_, analyze_, etc.). However, descriptions are inconsistent in depth, some are detailed (analyze_prompt: 180+ chars with trigger shortcuts), others minimal (get_settings, get_stats: <50 chars). Parameter descriptions are present but often generic ('프로젝트 ID'). Output schemas are not documented, LLMs cannot predict response structure. Error handling uses a generic safe() wrapper returning text, with no recovery guidance or error classification. No tool annotations (readOnlyHint/destructiveHint). Composition is reasonable (21 focused tools), but several tools conflate concerns (e.g., analyze_prompt handles local/api/deep modes, tagging, versioning, and project storage in one call).

• Document output schemas for all tools. Example: 'Returns {projects: [{id, name, promptCount, avgScore}], total: number}'. LLMs need to know response structure to chain calls.
• Replace generic error handling with actionable messages. Instead of 'Error: project not found', return 'Project "xyz" not found. Available projects: proj-a, proj-b. Call list_projects() to see all.'
• Add tool annotations to index.js. Use server.tool(..., {...}, {readOnlyHint: true}) for get_* tools and {destructiveHint: true} for create_/delete_/import_ tools.
• Rename hyphenated tools to underscores: quick_analyze, deep_analyze, project_report, code_review. Aligns with MCP naming convention and verb_noun pattern.
• Expand descriptions for get_settings and get_stats to 80-120 chars. Explain what each reveals and when to call it (e.g., 'Call before set_api_key to check current model and API status').