If two rounds of prompting haven't fixed a bug, a third won't either. The model doesn't have the context it's missing, and each round adds another locally-plausible change to a system that already doesn't cohere. At that point someone needs to read the code.
The pattern is recognisable. You describe the symptom, the model changes something, the symptom moves. You describe the new symptom. Four rounds later you have four modifications, each of which looks reasonable on its own, and a bug that has migrated rather than disappeared.
This happens because the model is reasoning about your description, not about your system. It cannot see the file that contradicts the fix it just wrote, and you have no particular reason to suspect that file either.
A few of these close the loop without anyone else involved. They're worth twenty minutes before asking for help.
A bug that can't be triggered on demand can't be confirmed fixed. Getting a reliable reproduction is often most of the diagnosis.
I trace one value from entry to failure rather than reading files top to bottom. The hop where it stops matching expectations is the bug.
A try/catch that hides the error is not a fix — it's the same bug with less visibility. If the real fix is bigger, I'll tell you that before doing it.
Generated code repeats its mistakes. Where one seam is wrong, the ones built the same way usually are too.
Yes, and it’s a large part of what I do. Generated code tends to be locally sensible but globally inconsistent, which is exactly the kind of problem that’s hard to prompt your way out of. I read it the same way I’d read any unfamiliar codebase.
Yes. If something is broken for real users right now, say so when you get in touch and I’ll treat it accordingly. I’ll want access to logs and the deployment, not just the code.
The service
The app half-works. You've described the bug to the AI six times, and each fix moves the problem somewhere else. At some point the loop stops being cheaper than reading the code.
A practical method for debugging AI-generated code when prompting has stopped helping: reproduce it, read the seams, check the assumptions, and fix the cause.
Read
8 min read
When Cursor can’t fix its own bug: use git to see what it changed, bisect to the breaking edit, and check the four places Cursor code usually fails.
Read
9 min read
Sometimes it needs an engineer.
Send the messy version. That’s usually the useful version.