Minesweeper MCP has complete input schemas for all 6 tools with proper JSON Schema structure (types, descriptions, constraints). Tool names follow verb_noun convention (create_board, reveal_cell, flag_cell, get_board, list_boards, delete_board). Descriptions are present and adequate (40-80 chars), explaining what each tool does. However, descriptions lack WHEN to use guidance and consequences of state-changing operations. Parameters have descriptions and constraints (min/max for dimensions). Output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance provided. No tool annotations (readOnlyHint, destructiveHint) despite clear risk classifications. Overall solid foundation but missing production-grade polish around output documentation and error guidance.

• Document output schemas for all tools. Example: create_board returns {boardId: string, width: number, height: number, mineCount: number, board: string[][]}. get_board returns {boardId: string, board: string[][], gameState: 'active'|'won'|'lost'}.
• Add state-change language to descriptions: 'Creates a new board (irreversible)' for create_board; 'Reveals a cell and updates board state' for reveal_cell; 'Permanently deletes the board' for delete_board.
• Implement tool annotations: add readOnlyHint: false to create_board, reveal_cell, flag_cell, delete_board; add destructiveHint: true to delete_board; add idempotentHint: true to get_board, list_boards.
• Add error handling with recovery guidance: 'Board not found. Call list_boards() to see available boards.' for invalid boardId; 'Invalid coordinates: x must be 0-{width-1}, y must be 0-{height-1}' for out-of-bounds.
• Add pagination to list_boards: accept optional limit (default 20, max 100) and offset parameters; return {boards: [...], total: number, hasMore: boolean}.