Separate correctness from presentation
Start with observable behavior: inputs, outputs, state changes, and error responses. Compare these with requirements written independently of the generated solution. A neatly named function can still calculate an incorrect total or authorize the wrong actor, so style checks belong after behavioral questions.
Illustrative scenario: a generated billing routine passes tests for whole-number quantities but rounds each line before applying a shared discount. Add decimal quantities, refunds, and currency rounding rules to the review. The issue is an incorrect business contract, not merely untidy code.
Inspect the cost of future changes
Look for duplicated business rules, unexplained abstractions, hidden global state, and modules that mix storage with policy decisions. Ask a maintainer to identify where a likely next change belongs. If the answer requires editing several unrelated files, investigate the coupling before adding more features.
Check whether new dependencies solve an actual need and whether errors carry enough context to diagnose failures. Prefer explicit boundaries around external services. Maintainability evidence can include a small follow-up change: how much unrelated behavior must be understood to modify one documented rule?
Build an evidence-based acceptance checklist
Require focused tests for the changed behavior, integration checks for affected boundaries, and review by someone familiar with the domain. Measure unresolved findings by severity and uncertainty. Counting generated lines, comments, or passing assertions alone provides little information about whether the feature is dependable.
Record remaining limitations in the pull request and backlog. Separate correctness defects from maintenance improvements and security findings so owners can prioritize them appropriately. A useful acceptance record explains what was tested, what was inspected, and which assumptions still need validation in a representative environment.
Working sheet
| Dimension | Review prompt | Evidence |
|---|---|---|
| Correctness | Which contract could this violate? | Independent expected outcomes |
| Maintainability | Where does the next rule change belong? | Focused modification walkthrough |
| Integration | Does the installed API match this call? | Version-specific build and integration test |
| Test quality | Would this catch the original defect? | Failing test on the unpatched revision |
Questions about how to assess AI-generated code quality
Is code coverage a quality score?
Coverage shows exercised code, not whether assertions are meaningful or requirements complete. Pair it with risk-focused review.
Should every generated abstraction be removed?
No. Keep abstractions that clarify real responsibilities; question those that add indirection without a demonstrated use.
How should I use the working sheet?
Download the CSV and replace the sample rows with evidence from your repository. Record a clear acceptance check and an owner for each item. Keep unresolved assumptions visible rather than treating the example as a completed assessment.
Sources and further reading
Examples are illustrative. They are not customer results, security guarantees, or a substitute for checking the documentation for your installed version.
