What's something AI coding assistants are surprisingly *bad* at?
I spent yesterday debugging a payment flow that an AI assistant confidently generated for me. The happy path? Flawless. The part where a user's card declines mid-transaction and we need to retry with exponential backoff while logging to three different services? The assistant completely missed it—and didn't even flag that it was skipping anything.
It got me thinking about where these tools actually excel versus where they consistently fall short. They're genuinely exceptional at scaffolding, at spinning up the obvious structure of a feature. But the moment you need to handle what shouldn't happen—race conditions, partial failures, the weird state your system ends up in at 2 AM—they seem to lose the plot entirely.
The frustrating part isn't that they can't solve these problems. It's that they sound equally confident about both the routine stuff and the edge cases. You have to actively know what you're looking for to catch the gaps.
I'm curious: what's something you've seen an AI coding assistant do really well that surprised you, and what's something it confidently botched that caught you off guard?
Replies
I’ve noticed the same with edge cases. The code can look perfectly reasonable until you ask what happens when three things fail at once.
For me, AI assistants are great at turning a rough idea working code, but I still slow down around payments, auth, and anything involving retries.
Two things from the last quarter of solo building that really stood out to me.
AI has been great at the boring mechanical stuff. Things like “rename X to Y everywhere it appears, update the tests, etc.” A 90-minute grep-and-pray job is now maybe 10 minutes with a proper review. Same with things like adding schema.org markup across a bunch of dynamic route templates. Basically, if it's pattern-matching, it's usually great.
Where it got me was timezone stuff.
My daily puzzle rolls over at midnight America/New_York, and I had AI generate the date logic a few different times. It looked right, passed my initial testing, and was still quietly wrong for users in other timezones. It kept assuming UTC without really making that assumption obvious.
That's probably the thing I've noticed most: AI isn't necessarily bad at a particular type of problem. It's bad at telling you which assumptions it made while solving it.