The first instinct, when a model says something odd, is to correct it. Ask again. Steer harder. Get it back on the rails. I've done that plenty — especially when I'm on a deadline and weirdness feels like friction.
Lately I've been trying the opposite. When the reply goes sideways — a metaphor that doesn't quite land, a connection I didn't ask for — I sit with it for a minute. Sometimes it's noise. Sometimes it's the most interesting thing in the room.
Safe outputs are easy to recognize: balanced, hedged, ready for a slide deck. The weird stuff is where the tool stops performing competence and starts accidentally thinking adjacent to you. I don't always keep those lines. But I've learned not to sand them away on contact. A little strangeness is often the start of a better second draft than another round of "make it clearer."