# The code review loop

Agents write a lot of code quickly. The bottleneck moves from typing to
judgment: deciding what's actually right. A review loop makes that judgment
repeatable instead of ad hoc.

The loop is small. After each change, ask for a review that leads with the
biggest risk, read it critically, and decide what lands. The point isn't to
outsource the call — it's to surface what you'd otherwise skim past.


**Review the working changes**

```
Review the current diff for correctness and clarity.

Lead with the highest-risk issue. For each finding, name the file and the
reason it matters; skip style nits unless they hide a bug. End with a one
line verdict: ready to land, or what to fix first.

```


Run it, read the verdict, and either land the change or send back the one thing
worth fixing first. That's the whole practice.