Strong schema definitions with Zod validation across all 27 tools. All tools have descriptions (10-100 chars) and input schemas with typed parameters. Naming follows verb_noun convention consistently. However, descriptions are often terse (many under 50 chars), lacking context on WHEN to use tools or dependencies between them. Output schemas are not documented, LLMs cannot predict response structure. Error handling maps to MCP error codes but lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk stratification in metadata.

• Document output schemas for all tools. For list_* tools, specify: fields returned, pagination (limit/offset/continuation), total count. Example: 'Returns {articles: [{id, title, content, feedId, timestamp}], continuation: string|null}'.
• Expand descriptions to 80-150 chars. Add WHEN context: 'List articles from FreshRSS. Use this to discover unread items; for starred articles, use list_starred_article_ids instead. Supports filtering by feed, category, label, or predefined states (reading-list, starred, read, unread).'
• Add tool annotations to tool definitions. Mark read-only tools with readOnlyHint=true, destructive tools (unsubscribe, delete_*) with destructiveHint=true, and idempotent tools with idempotentHint=true.
• Enhance error messages with recovery guidance. Instead of 'FreshRSS authentication failed', return 'Authentication failed. Verify FRESHRSS_URL, FRESHRSS_USERNAME, FRESHRSS_API_PASSWORD are set correctly and API is enabled in FreshRSS settings.'
• Document parameter formats and constraints in descriptions. E.g., 'feedId: numeric feed identifier (e.g., 123)' or 'streamId: greader format (feed/123, user/-/label/FolderName, user/-/state/com.google/reading-list)'.