The server has 4 well-defined tools with proper schema registration using mcp-go. Tool names follow verb_noun convention (deck_names, add_note, add_notes, find_notes). All tools have descriptions and documented input parameters with types. However, there are notable gaps: (1) Output schemas are not documented, handlers return plain text via mcp.NewToolResultText(), losing structure that LLMs need for downstream composition. (2) Parameter descriptions are minimal (10-50 chars), lacking guidance on constraints, formats, and use cases. (3) No error recovery guidance, errors return generic failure messages without actionable recovery hints. (4) No idempotency guarantees documented for write operations. (5) The add_notes tool expects a JSON string parameter rather than a structured array, forcing the LLM to serialize and the tool to deserialize, poor UX. Overall, the toolset is functional but lacks production-grade polish in output structuring, error messaging, and parameter ergonomics.

• Restructure output to return JSON objects with typed fields. E.g., add_note should return {"id": 12345, "deck": "Default", "model": "Basic"} so the LLM can extract the note_id and pass it to future operations. Update tool descriptions to document the return schema.
• Expand parameter descriptions to 50 - 150 characters, including format, constraints, and examples. E.g., 'Front content of the note (plain text or HTML, max 5000 chars). The question or prompt shown first when reviewing the card.'
• Add recovery guidance to error messages. E.g., 'Deck not found. Available decks: [Default, Spanish, Math]. Call deck_names() to list all decks.' This enables self-correction without an extra LLM round-trip.
• Refactor add_notes to accept a structured array parameter (mcp.WithArray('notes', ...)) instead of a JSON string. This makes the schema self-documenting and eliminates serialization overhead.
• Document idempotency guarantees. Add a note in tool descriptions: 'This tool is idempotent, calling it multiple times with identical parameters will not create duplicate notes (Anki deduplicates by content hash).' If AnkiConnect does NOT guarantee this, implement client-side deduplication with a request ID.
• Add limit and offset parameters to find_notes with defaults (limit=50, offset=0). Update the description: 'Find notes using a search query. Returns up to 50 note IDs by default. Use offset to paginate through larger result sets.'
• Validate inputs early and return clear, actionable errors. When a required parameter is missing or malformed, return the exact constraint that was violated (e.g., 'deck_name must be a non-empty string; got empty string.').

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry