The bug that cost us three days was in code I'd never read, and I approved it twice
We shipped a render queue in about four hours with Claude Code. It worked fine. Three weeks later renders started failing at roughly 1 in 40, and it took three days to find, because the retry was firing against a stale job id. I had approved that diff twice without reading past the function signature.
That isn't an AI problem. It's that approving generated code feels like reading it, and it isn't. When I write something badly I remember roughly where the bad part is. When I approve something badly I remember nothing at all, so debugging starts from zero.
What I do now is type out anything touching money, queues or user data by hand, even when the agent drafted it first. Everything else gets generated and I accept I'll debug it blind. It's about 15% of the codebase and it's the only 15% I can reason about at 2am.
Where do you draw that line, or do you not draw one?
Replies
I’ve felt this too. If I approve AI-written code without building a mental model of the critical path, I’m basically borrowing debugging pain from the future. For me, anything touching data, queues, auth or money still deserves a proper read.
@abhishekpatel Borrowing debugging pain from the future is right, and the interest rate is the part that got me. Reading that diff properly was maybe twenty minutes of work. Finding it three weeks later was three days. The one I'd add to your list is anything with a retry in it, because retries only misbehave under conditions that don't exist on the day you approve them.