In this case I just used Claude directly with a PDF, yes. In some other contexts I use a harness. Two different scenarios. First, I gave Claude the full manuscript and asked it to analyze it across various metrics (plot, characters, theme, prose, etc). It doesn’t know I’m the author. Second, with memory between chats disabled, I started fresh and just gave it the first third. I asked it to analyze it and make predictions about what will happen. It was more accurate than most human beta readers were at predicting. I then gave it another chunk and asked it to update its predictions, and indeed it got even more accurate.