Single tool 'toolguard_scan' has a basic description (58 chars) and minimal input schema. The description lacks actionable context for LLM selection (no WHEN to use, no prerequisites, no output shape). Input parameter 'target' has a type but no format guidance (absolute vs relative path? directory only?). No output schema documented. No error handling guidance. Tool does one thing (scan), but definition quality is well below production baseline.

• Expand description to 100-150 chars: 'Scans a directory for code quality issues (TODO/FIXME/XXX markers) and returns findings with severity levels. Use this to audit code before deployment or identify technical debt. Requires read access to target directory.'
• Add format constraint to 'target' parameter: 'Absolute or relative filesystem path to a directory. Must be readable by the process. Relative paths are resolved from the current working directory.'
• Document output schema in description: 'Returns a JSON object with keys: tool (string), findings (array of {id, sev, where}), score (integer count of findings).'
• Add error handling examples: 'If target does not exist, returns error with message. If target is a file (not directory), returns error. If permission denied, returns partial results with accessible files.'
• Return structured object instead of JSON string. FastMCP should deserialize the response automatically; if not, wrap scan() output in a dict and let the framework serialize.