How do you review code when AI writes most of the first draft?
I’ve been using AI coding tools Claude and Codex more often lately, and I keep running into the same problem:
I explain the task clearly
The agent writes a big chunk of code
The code looks correct at first glance
I start reviewing it and realise I don’t fully know why some decisions were made
Now the “fast” part becomes slow because I have to reverse-engineer the changes
The issue isn’t that the AI writes bad code every time. Sometimes it writes useful code. The harder part is staying confident that I understand what changed, why it changed, and what might break later.
Curious how others are handling this:
Do you ask the agent to explain every change before applying it?
Do you review file by file manually?
Do you force smaller commits?
Do you write tests first and let the agent work around them?
Or do you just accept that reviewing AI code is now part of the workflow?
I’m trying to figure out if people have a clean process for this, or if everyone is still experimenting.
Replies