What Forty-Seven Revisions Actually Looks Like
So, the question is, why would it take dozens and dozens of iterations to finish a chapter if the AI is able to write 100k lines before you can fill your coffee up?
Well, I have this feeling sometimes, when talking to people about this, that the idea of an AI workflow in writing looks something like "Hey Claude, write me a book about vampires that will win a Pulitzer Prize." If you're reading this, pen in hand, waiting for me to get to that part then I should apologize early on, it's going to be a slightly longer road.
What AI can do is provide a singularly amazing platform for someone to explore ideas, even interact with their own ideas in a pseudo-conversational level. It does take some care, and we really must be honest about the limitations, but my feeling is the reward for this investment is worth it.
My approach to setting up an AI workflow is rooted in my experience using it for coding projects and personal research projects. I have not changed this for writing. I will be happy to provide details on that in another session if there is interest.
For this session it must suffice to say that from the beginning, I worked on the entire story, not one thread in isolation. The entire four-book story arc was outlined at once; all character bios and motivations, biological and evolutionary traits, hidden world mechanics, physical location descriptions. Everything. Each of these is constructed as a conversation within a single AI session specializing in that specific thread.
That makes up the base set of information required for the AI to have a conversation about the Beat Sheets. But that conversation, as thorough as it may be, is roughly where the "AI wrote a book" fantasy ends and the actual work begins.
The first draft came out at around forty thousand words. Structurally complete, fifteen chapters, beginning to end. It read like a capable summary of the book I wanted to write. The beats were correct. The voice was absent. Every chapter had the same emotional temperature, which is a polite way of saying none of them had any temperature at all.
It was hollow. That's the word I keep coming back to. Grammatically flawless, structurally sound, and hollow. The AI produces prose that is technically perfect and emotionally vacant, it doesn't know what a sixteen-year-old girl would actually say, how a sentence should land in the chest rather than the head, where a paragraph needs to breathe. That hollowness is the reason forty-seven revisions exist. Every single pass in this process follows the same loop: the AI generates or analyzes something, I read it, I find what's wrong or what's missing, I rewrite it in the voice I hear, and I send it back. Then we go again. The AI provides the scaffolding and I provide the building.
Version two introduced the voice, a first-person interiority where the protagonist manages her world through a private filing system, categorizing every sensation, every interaction, every unnamed thing that is happening to her body. That voice became the book's engine. It also created the book's biggest problem: a manuscript that was ninety-two percent interior monologue. Characters who were physically present in scenes went entire chapters without speaking. A classroom had a teacher in it, students around her, a boy she couldn't stop noticing three rows away…and nobody said anything for two thousand words. It was a prison not a story.
Then came the dialogue pass. We measured every chapter. The manuscript was running at 7.9 percent dialogue. For context, literary YA like Kazuo Ishiguro's Never Let Me Go runs around 18–22 percent. Stephenie Meyer's Twilight is closer to 25–30 percent. We were at less than eight. Some chapters, namely the ones that should have been the most dramatic in the book, were below three percent. Fixing this isn't about adding lines. It was about rethinking how scenes worked. We developed and applied an OIP cycle (Observation, Interaction, Processing) a rhythm where the protagonist observes, then someone speaks, then she processes what was said. The interior monologue didn't shrink; it just got given voices. Every major scene in the book was rewritten around this architecture. And when I say rewritten, I mean by me. The AI proposed the structure. I wrote the words.
Then the continuity errors surfaced. This is the part nobody warns you about when you're working iteratively. Characters were sitting at a school table on a Sunday. A friend's running joke contradicts itself across two chapters because Tuesday's punchline doesn't match Thursday's callback. Day counts that were emotionally satisfying, thirty-two days, thirty-nine days, turned out to land on Saturdays when the scenes required Mondays. These aren't the kinds of errors an AI catches on its own, because each scene is internally coherent. It's only when you read across chapters, tracking a calendar that exists in the world but not in any single paragraph, that things suddenly don't make sense. I found most of these myself. The AI confirmed them. We fixed each one, and each fix rippled into the next scene, which rippled into the chapter after that.
Then there was the epigraph architecture. Each chapter opens with a quote by real authors, real works, and a comment line beneath it, written in an unnamed voice. The quotes are arranged in a progression that runs the length of the book, and the comment lines are a second, hidden conversation that only becomes legible on a re-read. This sounds elegant in the abstract. In practice, it meant a poem that opened the book created eighteen lines of front matter before the reader met the protagonist. It got moved to the back. Next, a chapter designed without an epigraph ended up looking like an oversight rather than an intentional break. A quote assigned to one chapter created a resonance problem with a physical book that appears inside that same chapter. Every change to one epigraph affected three others. This took four dedicated sessions of the AI proposing possibilties, me reading them against the manuscript, deciding what worked, and rewriting the lines myself.
Then the mechanical passes…and there were many. A style audit where phrases that started as deliberate callbacks had replicated until they became tics: "without meaning to" appeared eleven times and was cut to five, each one now load-bearing. "The quality of someone who" went from twenty-four instances to twelve. "She had been" dropped from seven per chapter to four. Each cut required reading the surrounding paragraph to make sure the rhythm still held. Seven hundred and sixty-eight straight quotation marks that had crept in through round-trips between the word processor environment and the drafting environment. An expansion pass was applied to reach the sixty-five to seventy thousand word YA target, where every addition had to pass a test: is this new material doing the same kind of work as the prose around it, or a different kind? If different, it didn't belong.
Then the five-agent editorial simulation. This is something the AI is genuinely excellent at, and it deserves its own blog session. In brief: a simulated panel of five editorial voices (a commercial agent, a genre specialist, a literary quality reader, a series-engine analyst, and a literary fiction benchmark) each read the manuscript and gave structured feedback. They disagreed with each other. The commercial agent wanted more physical sensation in the romantic scenes. The literary benchmark wanted less. This produced an endless list of action items. And what made this work was the action items were diagnoses, not prescriptions. The AI told me what wasn't working. Then I would decide how to fix it and go write the fixes. It takes 1-3 sessions of that loop per chapter before we land at something nearly readable.
Then the filing system recalibration. The protagonist's interior filing system, the metaphor that drives the entire narrative voice, had been appearing eight to ten times per chapter. The target was three to four. But you can't just delete references to the central metaphor of your book. Each cut had to be replaced by the thing the filing system was substituting for: the actual observation, the actual failed response, the actual moment where the protagonist reaches for the drawer and the drawer isn't there.
At one point the series title changed. Which meant a full manuscript review — not find-and-replace but reading every instance and asking whether the new title carried the same weight in that specific sentence.
And then, after all of that, the final editing pass on each chapter was mine alone. Because even after forty-six rounds of collaborative work, the AI is still hollow. The words land correctly and don't sound like a person wrote them. A sentence is technically beautiful and not what this character would say. A paragraph has the right information and the wrong music. This pass has often been the most extensive of all, because the AI tends to favor certain phrases and literary constructions that don't fit the voice I hear when I read the story in my head.
So, when I say forty-seven revisions, I don't mean I wrote the same chapter forty-seven times. I mean that the chapter, and the manuscript around it, went through forty-seven different kinds of attention. Some were structural demolitions. Some were single-word fixes that took an hour to locate and thirty seconds to implement. Some were philosophical recalibrations where the AI and I spent an entire session debating whether a character's silence in a scene was ethically motivated or just narratively convenient, and the answer changed three lines of dialogue and nothing else.
What I hope to show, in being open and transparent here, is that what actually happened is something I don't think we have good language for yet; a process where I could hold the entire architecture in a shared space, argue with it, measure it, break it and rebuild it, at a pace and depth that would be impossible alone. A process which requires a second mind in the room, and the second mind needs to have read everything, remembered everything, and be willing to tell you that your favorite sentence is the one that needs to go.
That is what forty-seven revisions looks like. It's not automation. It's a collaborative and maddening conversation with a system that is very good at seeing the thing you're standing too close to, and completely incapable of caring about it the way you do.
Which is, it turns out, exactly the combination you need.