I used Claude to code something. The thing is, it happilly creates the code, which looks quite professionally, and is soooo positive about itself. Then you try to run it, which ofcourse doesn't work. Next you feed it the error messages and it very very happilly fixes those bugs, all while being very fond of itself. After a few rounds of that, the code actually runs and does something.
Now I can get that it doesn't work from the first try, ours won't be 100% correct either, but the mistakes it makes tend to be because it mixes information of different versions of libraries.
And why is the damn thing so fond of itself? Everything it does it find "perfect".