Writing code with an AI is a private conversation right up until you open a pull request. You know what you asked for, what you tried, and why the code ended up this way. The person reviewing it doesn’t have that conversation in their head.
Read the change properly yourself before you ask anyone else to review or merge it. Handing over code no human has read, to someone with none of your context, is dumping your work on them.
Producing got cheap and reading never did.
Before asking for review #
Read your own diff. All of it. Line by line, the way you’d read a stranger’s code. You’ll spot the leftover debug log or the unrelated edit in seconds because you know why it’s there. A reviewer has to stop and work that out. #
Run an AI review too. It will catch things you missed. Go through each finding and fix or reject it yourself, so you understand the change before anyone else approves it. #
Turn repeated corrections into rules. Every fix you make more than twice belongs in yourCLAUDE.md , so the model stops producing it and you stop cleaning it up.
What this does to the reviewer #
People adapt to what you send them. Once someone works out that your PRs are unread model output, they either slow down and audit every line, or they stop reading properly and approve on faith. You caused both and you don’t get to pick which one you get.
None of this is really about AI. Pasted snippets and unread boilerplate existed long before it. Models just made it fast to produce a lot of code you haven’t thought about.