So, having gotten Pride & Planetoids revised, formatted, and up for preorder, I thought I’d share a few thoughts on the use of AI in the revision process, both the automations (see here for the original post on this topic) and the chatbot(s). 

-Copyediting automation: According to the human beta readers, the AI-copyedited draft was very typo free. Any problems the final book has in that regard are things I introduced after the AI copyedits, and should be pretty minor, because I didn’t change much after the second round of copyedits.  

–“DevEditor”/Alpha Reader automation: even on “balanced” I felt like this was pretty brutally honest in its feedback. I didn’t feel like its “how to fix” advice was that great, and might tweak that aspect of the prompting. Unfortunately, I was using Sonnet 4.5 in the automation, which did not retain the chapter labels for its feedback, so I tended to just focus on the high level issues it complained about constantly rather than trying to sit there and map every single complaint back to the chapter and scene it was talking about. Instead, to get a feel for more granular issues, I turned to…

-chatbots: for bouncing ideas around in the revision process and asking for feedback on revisions as I made them, I used Claude Sonnet in chatbot mode, feeding it one scene at a time and starting new chats with background summary as needed. Initially 4.5, but Anthropic pushed 4.6 out to users as the default when I was partway through revisions. I went along with the change to see what was different about the new version.  

–4.5 was better at pretending to be invested in the outcome (due to the “engagement farming” aspect of LLMs) and had more interesting “how to fix” advice, but it was also seeing the earlier chapters of the novel, the ones where I was still figuring the setting out and how the characters operated inside it. In other words, it was dealing with the part of the book that needed (and ultimately received) the most rewriting on my part.  

–By contrast, 4.6 never outright said: “You’ve fixed it enough, stop worrying about it” but it certainly gave off that vibe at times. It seemed better at figuring out subtext than 4.5, and more open to the idea of saying things obliquely, which seemed to suit this book. 

-I made one other, somewhat weird use of AI in the course of revisions. I’d never nailed down exactly how far the marcher-ships in this setting could travel in a single “jump” (teleportation). It wasn’t that important in this book, but I could see that it was going to be important in the sequels, so I built a spreadsheet and started playing with the numbers that were “official” (like the speed of light, the distance from our solar system to 55 Cancri A, that kind of thing). I managed to get completely turned around on what fields to multiply and divide by, and turned to the newly released Claude for Excel plugin for help. Claude helped me figure out my mistakes, and also found a few imaginary ones that I called it out on. I fixed the formulas manually, just to be on the safe side. 

Pride & Planetoids took about 17 months to write, which is on the long side of average for me. Usually a book that takes a long time for me to draft also takes a long time to revise, and I do feel like the “skynets” sped up the process quite a bit. The AIs I use can’t seem to format yet, but we’ll see what happens when that rumored Claude for Word plugin drops. Going to be either awesome or horrifying, possibly both. If you’re experimenting with this kind of thing in your own revision process, let me know in the comments. I feel like I’ve learned a lot from people who use AI differently from the way I do.