I've been using Claude Code *extensively* for the last six weeks. It tests the ever-loving-daylights out of everything.
You ask it for X. It writes three tests to make sure X is happening before it tries to do X. It writes X. It notices a problem you didn't, thinks of a special case you didn't, and adds two tests for those. It fixes everything, then runs all five tests. Then it runs ALL SIX HUNDRED TESTS IT HAS ACCUMULATED SO FAR ON THIS PROJECT. It then reports on all of this. If any test fails it doesn't stop, it keeps working until it either needs to ask you a question or al tests pass. If something looks funny one time it will test for that every time it changes anything forever.
Of course, whatever you're working on has to interface with a lot of other things and you need wider scale testing as well as local testing, but that's true of human coding, too.
It's a tool. All machines are amplifiers. If you are lazy, if you are impatient, if you are under informed, you're going to get bad output, but you were going to get that bad output anyway, weren't you?
So the haters are going to say it's vibe coding, and that's hard to disprove. And the proponents of AI are going to say people didn't give the models enough info about the cope, and that's just No-true-scotsman.
But I am producing more and better code than ever. I have never accomplished so much so fast, and as more-or-less a hobbyist I never had access to testing like this. And it knows about security issues I don't, and at least one time has told me about a weakness I did not suspect at all.
There will always be problems with software, but AI should make them easier to find. It's probably the only way to practically test truly huge systems.
But there's no disproving the allegation that vibe coding is causing the problems, since 90% of code is machine-written now.
So it's just popcorn time…