I've had an LLM wired into my shell for a few weeks now. Type a plain-English description, get a command back, hit enter or edit it first. For the boring stuff it's genuinely good. I no longer go and re-read the find man page to remember the incantation for "files older than thirty days, bigger than 100MB". It just produces it, correctly, and I get on with my day.
The problem is that it produces the wrong thing with exactly the same confidence as the right thing. It once handed me a tar invocation that, had I run it, would have extracted over a directory I cared about rather than into a fresh one. The flags were plausible. The explanation was plausible. It was wrong. Another time it suggested a git command that would have rewritten history on a shared branch, cheerfully, as though that were a normal Tuesday.
Neither bit me, because I read the command before running it, which is the entire discipline. The tool is a very fast junior who has read everything and remembers nothing about my system. So I treat its output the way I'd treat a snippet off the internet: useful starting point, never a finished answer. The day I stop reading the command before I run it is the day it deletes something I'll miss. So far I'm still reading.