I do wonder if my software engineering experience is having a negative effect to the effectiveness of these agents. After a couple of decades in the field, picking up the lessons and scars of things going right and wrong, it seems natural to get a sense of what;s possible, or how something should be done. It could be that I’m just imagining it, but I feel these boundaries are coming through in the prompts I write. I’d research an approach and tell the agent in detail how to do it. Or I would leave certain behaviours out, or explicitly tell the agent not to do something, with the belief that asking for it would be impossible. Really keeping the reins quite tight.

This is sometimes warranted, when I need to know the code inside out. But for silly little vibe-codeing sessions, is it really helping? Wouldn’t it be better to loosen the reins a little and let the agent find these boundaries?