I've been experimenting a bit today with letting models reason for creative tasks, rationale being that it might help with keeping track of details and prompt adherence. And predictably, the wall I'm running into is that they all want to draft, check, refine, revise, "um actually...", and draft again before actually giving the output. This is hugely wasteful no matter the reply length, but it's especially bad if you want more than a paragraph or two. Obviously I'm not the first pers...