Three tools with clear names and documented schemas, but descriptions lack depth and parameter documentation is incomplete. 'observe' and 'inspect' have generic descriptions under 50 chars; 'act' lacks clarity on state management. Parameters like 'mode', 'kind', and 'args' have minimal guidance. No output schema documented. Error handling returns JSON but lacks recovery guidance. Tool names follow verb_noun pattern (positive), but descriptions do not explain WHEN to use each tool or dependencies between them.

• Expand tool descriptions to 100-200 chars. Example: 'observe: Fetch current game state in delta (smallest), compact (cached), or full (raw) mode. Use delta for token efficiency; full for complete state. Returns state snapshot, definition changes, and warnings.'
• Document output schema for each tool. Example: 'observe returns {state: {...}, defs?: {...}, warning?: string, error?: string}. Use state for game logic; defs for decoding; warning for alerts.'
• Replace 'args: object' with typed parameters or an enum of action types. Example: 'action: enum[play_card, end_turn, ...], card_index?: int, target?: string' so LLM knows valid combinations.
• Add enum constraint to 'kind' parameter in inspect(). Change to: 'kind: enum[map, state, wiki, profile, profiles, compendium, metrics]' in schema.
• Enhance error messages with recovery hints. Example: 'Cannot connect to STS2MCP. Ensure the game is running and the mod is enabled. Retry in 5 seconds or check game logs.'