Four tools with complete schemas and descriptions, but descriptions are verbose (200-300 chars) and lack LLM-optimized structure. Parameter descriptions are present but generic. No error handling guidance, no output schema documentation, no tool annotations. search and search_by_local_image have good semantic detail; image_info and download_images lack context on when to use them. No idempotency or retry guidance documented.

• Add explicit output schema documentation for each tool. E.g., search() returns {count: int, images: [{id, size, thumb, full, title?, caption?, source?, post_url?, original_url?, nsfw?}]}. This enables LLMs to plan downstream calls.
• Rewrite descriptions as action statements with context. E.g., image_info: 'Retrieve full metadata (dimensions, title, source, URLs, tags) for images by ID. Call after search() to get details before downloading or sharing.' Current: 'Full metadata for specific image ids...' is a list, not an action.
• Add error recovery guidance. E.g., search_by_local_image error 'File not found' → 'File not found: {path}. Verify the path exists and is readable. If unsure, use search() with a text description instead.' This guides LLM retry logic.
• Declare tool annotations via FastMCP metadata. Mark search() and image_info() with readOnlyHint=true. Mark search_by_local_image() and download_images() with destructiveHint=true (file I/O). This enables agents to reason about safety and retry semantics.
• Document parameter constraints explicitly. Add to search(n): 'number of results to return (1 - 100, default 30)'. Add to search(text): 'semantic description for CLIP similarity (e.g., "moody brutalist concrete at dusk"), keyword search does not work.' This prevents LLM hallucination of invalid values.