What reviews actually catch determines what they should look at. Logic errors, unhandled edge cases, design drift from the surrounding system and missing tests all reward human attention. Formatting, naming style and import order do not — they belong to linters and formatters, and spending review time on them consumes the scarcest resource in the process.
Size dominates everything else. Defect detection falls sharply once a change passes a few hundred lines, because reviewers stop reading and start skimming; a thousand-line pull request gets approved on faith. The discipline that makes review work is unglamorous: small, focused changes; a description that states intent rather than restating the diff; and review latency measured in hours, because review wait time is cycle time and stale pull requests rot.
AI coding tools have shifted the balance rather than removed the need. LLM review assistants work well as a first pass on mechanical issues, which frees human reviewers for design and intent — the questions of whether this change should exist and whether it fits the system. The harder problem is volume: generated code arrives faster than humans generate it, which makes review the bottleneck and raises the cost of the rubber stamp. The standard has to stay where it always was — nobody merges code they could not defend in an incident review.
High-risk areas deserve more than the default process. Authentication changes, data migrations, payment logic and permission models warrant reviewers with specific context — sometimes a required second approval from the owning team — because the cost asymmetry is extreme: ten minutes of specialist attention against an incident that outlives the quarter. Checklists earn their place here, not as bureaucracy but as the guarantee that the obvious questions get asked every single time, including the tired Friday afternoon ones.
Review is also the main channel through which system knowledge spreads. The reviewer learns how parts of the codebase they do not own actually work; the author learns the conventions and the reasoning behind them; newer engineers grow faster through reviewed work than through any training material. Teams that optimise review purely for defect capture and speed miss that they are also running — or neglecting — their most effective mechanism for keeping more than one person able to maintain any given system.
Codazz builds this in production — QA & Testing.