How to review AI-generated code without losing ownership of your project

The Core Question

AI can finish a feature faster than we can absorb its assumptions. The practical review question is not only "does this work?" but "could I explain why it works, where it can fail, and what it changes tomorrow?"

A Narrow-to-Broad Review Loop

A useful review loop starts narrow and expands only when needed. Begin with the selected expression, then inspect its function, file, dependencies, git change, and the project decision behind it. At each level, write down the one concept that is still unclear and test yourself against the actual code rather than a generic tutorial. This matters most when AI has produced code you need to maintain.

A passing test is necessary, but it does not prove you understand the boundary conditions, the data flow, or the coupling introduced by a change.

Unvibe's Learning Loop

I’m building Unvibe around this learning loop: select code, choose explanation depth, ask follow-ups, and retain concepts worth revisiting. The aim is not to outsource ownership again; it is to use the code already in the project as the curriculum.

What does your review process look like after Cursor, Claude, or ChatGPT generates a non-trivial change?