Server has 9 tools with explicit schemas and descriptions in Korean. Naming follows verb_noun pattern (search_, get_, recommend_, build_, generate_, optimize_, create_, clear_). Schemas use Zod with proper types, enums, and constraints. However, descriptions are verbose (200-400 chars) and include example values that LLMs may reuse literally. Parameter descriptions are present but inconsistent in depth. Output schemas are not formally documented, responses are inferred from code. Error handling is basic (try/catch with generic messages). No tool annotations (readOnlyHint/destructiveHint). One tool (clear_cache) is marked WRITE but lacks confirmation/dry-run pattern.

• Define formal output schemas for all tools. Document return types as JSON Schema (e.g., { type: 'object', properties: { items: { type: 'array', items: { type: 'object', properties: { goods_name, price, shop_url, ... } } }, total_count, has_more } }). This enables LLMs to extract fields reliably and chain tools.
• Replace example values in descriptions with enum constraints or regex patterns. E.g., instead of 'sort: relevance(관련도), popular(인기/판매량), price_low(낮은가격), price_high(높은가격), newest(최신)', use enum: ['relevance', 'popular', 'price_low', 'price_high', 'newest'] in schema and reference it in description.
• Add tool annotations to all tools. Mark read-only tools with readOnlyHint: true (search_products, get_best_sellers, get_product_details, recommend_for_lesson, build_classroom_kit, generate_semester_preparation_list, optimize_within_budget, create_shopping_list_from_text). Mark clear_cache with destructiveHint: true.
• Implement confirmation_request pattern for clear_cache. Add a 'confirm' boolean parameter (default false) and require explicit confirmation before wiping cache. Return a dry-run summary first.
• Enhance error messages with recovery guidance. E.g., 'No results for "query". Try: (1) broader keywords like "교구" or "스티커", (2) remove grade filter, (3) increase budget. Available categories: 학급운영, 보상체계, 과학실험.'.