Single tool 'run_claude' has a basic description but lacks critical production-grade patterns. The tool name 'run_claude' is generic and does not follow verb_noun conventions (should be 'execute_claude_command' or similar). Input schema is present with types and descriptions, but output is unstructured plain text with no documented schema. No error classification, no recovery guidance, no idempotency markers. The tool accepts a dangerous 'skip_permissions' parameter that defaults to true, creating a security risk. No pagination, no structured output, no tool annotations.

• Rename tool to 'execute_claude_command' to follow verb_noun convention and clarify intent.
• Return structured JSON: {status: 'success'|'error', output: string, stderr: string, exit_code: int} instead of plain text.
• Change 'skip_permissions' default from true to false. Require explicit user/agent authorization for dangerous operations.
• Add input validation: sanitize prompt parameter against command injection; validate verbose and skip_permissions as booleans.
• Implement error classification: return {status: 'error', code: 'not_found'|'timeout'|'permission_denied'|'invalid_input', message: '...', recovery_hint: '...'} with actionable guidance.