Delivery & maintenance
Review AI-assisted code by following the behavior
Review AI-assisted code for the behavior it introduces. Trace data, permissions, failures, and tests so a plausible implementation becomes a maintainable change.
An AI coding assistant can produce a large, convincing change quickly. The review still needs to establish that the change solves the intended problem and fits the application. Imagine a hypothetical customer portal receiving a generated document-download feature. The button and handler may look complete while the endpoint forgets to check which customer owns the document. A useful review follows the behavior from user action to stored data and back.
Start with the concrete requirement
Read the problem statement and expected outcome before reading individual lines. Identify the important negative cases: a missing document, an expired session, and a user requesting another customer's file. Ask the author to keep the change focused enough to review. If the patch includes unrelated refactoring or styling, separating that work can make the functional change easier to understand and the release easier to diagnose.
Trace the critical path
Follow the request through routing, authentication, authorization, data access, response handling, and the interface. Look for assumptions at each boundary. In the download example, hiding a button is not a substitute for checking access at the endpoint. Inspect failure paths and concurrency where relevant. Google's engineering review guidance emphasizes design, functionality, and complexity; those questions remain useful regardless of whether a person or an AI system drafted the code.
Read the tests as claims
A test is an assertion about behavior, not automatic proof of correctness. Check whether it would fail if the important protection were removed. Watch for generated tests that reproduce the same mistaken assumption as the implementation. Prefer realistic boundary cases over broad snapshots that are easy to approve without understanding. For a user-facing feature, inspect the actual interface or a representative demonstration when the diff cannot show the interaction clearly.
Leave a maintainable explanation
Require clear names and a short explanation for non-obvious decisions. Remove speculative abstractions that do not serve the current task. Verify new APIs against the project's actual dependency versions and official documentation. Summarize what was tested and any remaining limitation in the change request. The reviewer should be able to describe why the implementation works without relying on the AI assistant's confident narrative or the fact that the code compiles.
