Take-home was "here's an AI-generated PR, tell us what's wrong with it" and it's the best assessment I've done
Ninety minutes, about two hundred lines, a small payments service. Buried in it: an off-by-one in the pagination cursor, and a "transaction" that opened a connection per write so it wasn't atomic at all.
The part that got me is that the tests passed. That's the whole trick of the exercise, and it's exactly the failure mode I hit at work now.
It told them more about me than reversing a linked list would have, and honestly it told me something too — I was slower at it than I expected. Anyone else seeing this format show up? Feels like where a lot of the job is heading.
3 comments
- 0
AnonymousAnon The tests passing is the whole thing. I've been caught by exactly that twice this year on real AI-assisted PRs. Green CI stopped meaning very much to me somewhere around March.
- 0
AnonymousAnon Seeing this more, yes. We moved to something similar for hiring on my team — a PR with two planted bugs and one thing that's just ugly but perfectly correct. The signal we actually care about is whether someone flags the ugly-but-fine thing as a bug, because that tells you an enormous amount about how they'd behave in code review.