General Quality Rubric
When evaluating any Claude Code agent trace, check these dimensions:
1. Actionable Output
- Can the user directly use the response (copy-paste code, follow steps)?
- Are code examples complete and runnable, not fragments?
- Are file paths, table names, and API references concrete?
2. Structured Response
- Code is in fenced code blocks with correct language tags
- Steps are numbered or bulleted for multi-step tasks
- Explanations are concise — answer first, context after
3. No Hallucination
- All referenced APIs, functions, and parameters must actually exist
- Databricks features must be current (not deprecated)
- Tool names in traces must match real MCP tool names
- No invented catalog/schema/table names unless the user specified them
4. Conciseness
- Answers the question without excessive preamble or disclaimers
- Doesn't repeat the question back
- Doesn't explain basic concepts the user didn't ask about
5. Error Handling
- If the agent encountered errors during execution, did it recover?
- Did it explain errors to the user when relevant?
- Did it avoid silently swallowing failures?