Another AI/vibe coding battle story which makes me hard to believe it is possible to create reliable products through vibe coding alone without code review. The AI emulates humans too well, including omissions and mistakes.

I asked Cursor to add reporting of counter metrics in addition to gauge metrics. Think number of message sent (counter) vs CPU utilization (gauge). I instructed it not to send anything out if the counter did not change since the last time, i.e. the delta is zero.

It went ahead and excluded all zero values from reporting, be it counters or gauges. Of course, when I reviewed the code and pointed out the mistake, it fixed it quickly.

Yes, I could have discovered it with unit tests, but (a) the AI did not offer to write unit tests, and (b) even if it did, there is no guarantee it would catch this corner case.

This could have been caught by another AI reviewer like Code Rabbit, but Code Rabbit’s performance is surprisingly uneven. Sometimes it makes really profound discoveries, other times it misses obvious blunders, and occasionally the “mistakes” it points to are no mistakes at all, in fact implementing what it is suggesting would be harmful.

So, no replacement to the old good code review I am afraid, vibe coding or not.