The server exposes a single tool, convert_url, with a clear action verb and reasonable description. The tool has a properly structured input schema with four parameters, all typed and described. However, several definition quality gaps prevent a higher score: (1) The output schema is completely undocumented, the response structure is not specified anywhere in the visible code, forcing LLMs to infer what fields will be returned. (2) Parameter descriptions, while present, are minimal (10 - 30 chars) and lack actionable detail, e.g., 'CSS selector to extract specific content' doesn't explain the expected format or selector syntax, and 'Force browser rendering for JS-heavy pages' doesn't explain when this is necessary or its performance implications. (3) No error handling guidance is visible, the tool provides no recovery hints if the URL is invalid, inaccessible, or malformed. (4) The tool description (92 chars) is adequate but generic; it doesn't explain which output format to prefer or when to use selector vs. force_browser. (5) No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present, though this tool is read-only and idempotent and should declare both. The implementation delegates to convertUrl() in convert.js, which is not shown, so the actual output shape and error behavior are opaque. Baseline comparison: tool description length is below the p10 (34 chars minimum), parameter descriptions lack the 50 - 100 char range typical of A+ tools, and there is no documented return schema (baseline: 100% of A+ tools document return types).

• Document the output schema explicitly. Add a 'Returns' section to the tool description or a response schema block specifying: (1) for format='markdown': {content: string, title: string?, metadata?: object}; (2) for format='json': {content: object, title?: string}; etc. This allows LLMs to plan what data will be available and avoid invalid field lookups.
• Expand parameter descriptions to 50 - 100 characters with specific constraints. Example for selector: 'CSS 3 selector (e.g., .article-body, #main-content) to extract specific content. If no match, returns full page. Ignored if force_browser=false.' Example for force_browser: 'Force Puppeteer rendering for JavaScript-heavy pages (slower, ~5 - 10s latency). Required for SPAs. Overrides default heuristics.'
• Add tool annotations. Modify the server.tool() call to include hints: readOnlyHint=true and idempotentHint=true. This signals to LLMs that the tool is safe to call multiple times and will not modify remote state.
• Add error handling documentation in the tool description. Append: 'Returns error with guidance if URL is invalid, unreachable, or content cannot be parsed. If selector matches no elements, returns full page content. If browser timeout occurs after 30s, retry with force_browser=false or simplify the selector.'
• Clarify the format parameter. Update the description: 'Output format: markdown (default, Markdown text), html (HTML), text (plain text), or json (structured data with content, title, metadata). Choose markdown for LLM ingestion, json for structured extraction.'