It used to be good at reformating things (GPT-3 and earlier). I tried recently, it asked repetitive questions and forgot things. At one point it reformatted something poorly, and when I told it it was wrong, it said it was wrong because of me, it wouldn't be wrong if it didn't have to guess.
Agree one of the good use cases I've found is bung a wall of text into it, ask it to structure and tidy up, make coherent Yesterday with GPT5 it was really bad. Put things in the wrong places contextually, duplicated things.
And to be clear, it didn’t have to guess, I told it three clear instructions, it just only listened to one of them, and what it claimed it guessed was because it just didn’t remember two of the instructions. It was basically insulting me for its own inadequacies.
That really is wild! Adobe Firefly doesn't talk back, which allows me to see it as just tech. But I find that it can only really accept one or two things in a text prompt. And maybe not even that, unless it's something that's really common. It can't handle complexity. Too many things to juggle.