I've had an LLM bound to a keystroke in my shell for a while now. Ninety percent of the time it's quietly brilliant. It turns "rename every PNG in this tree to lowercase" into a find incantation I'd have spent five minutes squinting at, and it remembers the xargs -0 bit that I never do.
The other ten percent is why I'm writing this down. The failure mode isn't that it's wrong, it's that it's wrong with complete confidence and a plausible explanation. Last week it offered me a one-liner to "clean up merged branches" that included a git branch -D against a pattern broad enough to take out a branch I very much wanted. I ran it because it looked right. It looked right because the model is extremely good at producing things that look right. That is the whole trap.
So the rule I've settled on: it can suggest, I read every destructive command before it runs, and anything touching rm, git push --force, or a database gets pasted into a comment first so I have to consciously uncomment it. Friction, deliberately. The assistant is a very fast junior who has never once said "I'm not sure". I just have to remember that the not-sure is my job now.
It earns its keep. I'd not give it back. But I treat its confidence as a UI affordance, not a signal, and I'm slightly slower and a good deal less burned for it.