A checklist turns review from an open-ended request into a set of explicit conditions to inspect. This matters when evaluating AI-generated code: judges caught only 45–63% of known errors, while adding hints from a separate analytic checker raised coverage to 74%. section 8.2.3
Use targeted checks for recognizable failure patterns:
- Check for SQL injection vulnerabilities. section 8.2.3
- Verify that all inputs are validated. section 8.2.3
- Look for potential null pointer exceptions. section 8.2.3
- Identify hardcoded credentials and secrets. section 8.2.3
Checklists work best for pattern matching; mechanical checks such as documentation, naming consistency, and secret detection may need no judgment. They improve review focus but do not replace human reading or eliminate correlated model failures. section 8.2.3




