ScreenLook has 4 well-named tools with clear descriptions and Zod schemas. Tool names follow verb_noun convention (screenlook_capture, screenlook_devices, screenlook_diff, screenlook_doctor). Descriptions are concise (50 - 150 chars) and explain purpose clearly. However, parameter descriptions are sparse: device_id lacks detail on format/resolution behavior; wait_seconds has minimal guidance. Output schemas are not documented, LLMs cannot predict response structure. Error handling returns isError flag but lacks recovery guidance. No parameter validation rules stated (e.g., wait_seconds bounds). Missing tool annotations (readOnlyHint, idempotentHint). Overall solid foundation but gaps in parameter detail and output documentation prevent higher scores.

• Document output schema for each tool. Define response structure: { content: [{ type: 'image'|'text', data?: string, mimeType?: string, text?: string }], isError?: boolean }. Include field descriptions and example values.
• Expand parameter descriptions with format, bounds, and resolution logic. E.g., device_id: 'Device ID (UUID or name). If omitted, uses the only connected device; if multiple exist, lists them and returns error asking user to specify.'
• Add recovery guidance to error messages. E.g., 'Error: No devices found. Run screenlook_doctor to check prerequisites (adb, xcrun) or connect a device.'
• Declare tool annotations in schema. Add readOnlyHint: true and idempotentHint: true to all four tools so LLMs know they are safe to retry.
• Add timeout and cancellation support to screenlook_diff. Set a max wait_seconds (e.g., 300s) and document timeout behavior. Consider supporting MCP cancellation requests.