The Figma MCP server demonstrates good definition quality with two well-named, verb-first tools backed by detailed parameter schemas using Zod validation. Both tools have substantive descriptions (>100 chars). Tool names (get_figma_data, download_figma_images) clearly convey intent. Parameters are extensively described with regex patterns, enums, and format constraints. However, output schemas are not explicitly documented in the visible code, limiting visibility into what downstream tools receive. Error handling is present (validation rejection capture, path traversal checks) but recovery guidance in error messages is not visible in the provided code snippet. The download tool exhibits particularly sophisticated parameter engineering with conditional requirements (imageRef vs gifRef vs nodeId rendering), though this complexity could confuse LLMs without clearer conditional documentation. No security issues detected, credentials are not exposed as parameters.

• Document return types for both tools. Add a schema block to the tool registration showing: get_figma_data returns {content: Array<{type: string, text?: string, ...}>, ...}; download_figma_images returns {content: Array<{type: string, text: string}>, isError: boolean, ...}. This lets LLMs parse results correctly.
• Formalize parameter dependencies in download_figma_images. Add a paragraph in the tool description: 'Exactly one of imageRef, gifRef, or nodeId must be used to identify the render source. imageRef for raster fills, gifRef for animated GIFs, nodeId for vector SVGs.' Then update each param description to reference this.
• Add numeric constraint to get_figma_data depth parameter: 'OPTIONAL. Positive integer (1-10 recommended). Controls how many levels deep to traverse the node tree. Larger values increase response size and latency. Only use if explicitly requested by user.'
• Expose error message text in tool descriptions. E.g., 'If localPath traverses outside the image directory, returns: "Invalid path: ... Points outside image directory."' This prepares LLMs for the error before it occurs.
• Return structured metadata in both tool responses. download_figma_images should return {successCount, failureCount, downloads: [...], ...} to let LLMs understand partial success and plan retries.