I think it's very useful but the hype promises so much more than it delivers. And a lot of the proponents are all in on the hype it gets annoying.

I use claude to write a design, review the design, turn that into an implementation plan, spend 2-3 turns reviewing that, but still when that is turned into code it misses things or creates helpers that's not actually used or... It creates massive files and unless I explicitly tell it to it never refactors them. It often just silences errors and warnings instead of actually fixing the problem.

It saves a lot of time, and I'm building things I couldn't have on my own. But it makes a lot of mistakes, it's far far from one shots which the hype keep going on and on about. It's tricky to put firm limits on what it does. A lot of the mistakes I catch because I've spent 15 years without an agent and sometimes it's just "hm, this smells weird" and I begin digging. I worry about the next generation.

For me the mental framing of "It's all hallucinations, some of those hallucinations are useful" is helpful to keep frustration in check as I ask it to review the same implementation plan for the 4th time and it turns up different issues because the input was slightly different, or review the output code and see allow(dead_code) despite my claude.md forbidding it.

This is very much my experience and process as well. Lately I've found 4.7/4.8 telling me not to fix problems lol: "outside the scope of this PR," "not caused by your changes," "safe to ignore" etc. Robot, I decide what to fix, when, and how!

It hates following project conventions, loves mocking so much out of tests that it fails to test anything at all, and yeah constantly leaving dead code around.

And yet it's still very useful. It's just nowhere close to as good as it's hyped, and yet there are professional developers who are using it as though it is.