7 Taking stock
I've now reviewed all of round 3 (6): eight sub-chapters and thirty designs.
I'm not yet happy with the hands-off approach. For round 3, Claude did the research, briefed ChatGPT, built every model in FreeCAD and wrote it all up, with very little input from me. Going through it, I found that Claude didn't make accurate representations of the concepts. The FreeCAD models are meant to turn ChatGPT's ideas into real designs, and too often they don't represent the concept they came from.
So that's what I'm going to work on next.