USB MCP server exposes 4 read-only tools with clear verb-noun naming (usb_search, usb_get_skill, usb_audit_skill, usb_render_install). All tools have descriptions (194 - 250 chars, within baseline 34 - 392 range) and complete JSON Schema input definitions with typed parameters. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Parameter descriptions are present but minimal (10 - 30 chars); several lack format/constraint guidance. No error handling guidance visible. No tool annotations (readOnlyHint, etc.). Composition is sound: each tool has one responsibility, and naming clearly distinguishes them. Security is adequate for read-only operations.

• Document output schemas for all 4 tools. Example for usb_search: {type: 'object', properties: {results: {type: 'array', items: {type: 'object', properties: {slug, name, category, description}}}, total: {type: 'number'}, limit: {type: 'number'}}}.
• Convert 'target' parameter to enum in usb_search, usb_get_skill, usb_render_install. Current description 'Runtime target (optional)' does not list valid values; LLMs will hallucinate.
• Expand parameter descriptions to 50 - 100 chars with format/constraint guidance. E.g., 'query: Free-text search term (1 - 200 chars, matches skill slug, name, category, description)' instead of 'Search term'.
• Add error handling guidance: 'If skill not found, returns 404 with available skills in similar category. If target unsupported, returns 400 with list of valid targets.'
• Add tool annotations: readOnlyHint=true and idempotentHint=true to all 4 tools to signal safety to MCP clients.