So far, using Claude, I've produced two error reports of it, and absolute 0 actual, usable content.
Okay, not 0 usable content per se; it did correct the README of one of the project crates.
One time, trying to figure out a compiler segfault in a pipeline, it did suggest two changes to reduce the memory usage, none of them worked, and then decided it needed to change two different repositories, which I told it not to do 'cause those were used by other compilers, and then decided to change the compiler in question, which I, again, told it not to do.
The second time I asked for suggestions improving a generated markdown, it did suggest changing the layout a bit, I saw okay, then it did removed one of the fields, then added it back in the wrong format, then ignored the original format had a lot of decimal places displayed.
Both cases, I had to told Claude to fuck off and revert all its stupid changes (yes, exactly like that).
So far, my experience with AIs is: I'm writing readable READMEs explaining what each crate does, its constrains and such.
The only thing AI is helping me so far is forcing me to write proper documentation.