If an AI can write a hundred thousand words before you've refilled your coffee, why would a single chapter take dozens — by now hundreds — of passes to finish?
When I describe this to people, I get the sense they imagine the workflow as "Hey Claude, write me a vampire book that wins a Pulitzer." If that's what you're waiting for, I should apologise early: it's a longer road.
What AI does provide is a platform for interacting with your own ideas at a conversational level. It takes care, and you have to be honest about the limitations, but the return on the investment has been worth it.
The setup
My approach comes from using AI for coding and research projects, and I haven't changed it for fiction. From the start I worked on the whole story, not one thread: the full four-book arc, character bios and motivations, the biology and evolutionary mechanics, the hidden-world rules, the locations. Each thread was built in its own dedicated session. That corpus is what the AI needs to have a useful conversation about beat sheets.
And that is roughly where the "AI wrote a book" fantasy ends and the work begins.
The hollow first draft
The first draft came in around forty thousand words. Fifteen chapters, structurally complete, beats correct. It read like a capable summary of the book I wanted to write. Every chapter had the same emotional temperature, which is a polite way of saying none had any.
Hollow is the word I keep returning to. Grammatically flawless, structurally sound, and vacant. The model doesn't know what a sixteen-year-old would actually say, how a sentence lands in the chest instead of the head, where a paragraph needs to breathe. Every pass since has followed the same loop: the AI generates or analyses, I read, I find what's wrong or missing, I rewrite in the voice I hear, I send it back. The AI provides scaffolding. I provide the building.
The passes
Voice. Version two introduced the engine of the book: a first-person interiority in which Vess, the narrator, manages her world through a private filing system, categorising every sensation and every unnamed thing happening to her body. It also created the book's biggest problem — a manuscript that was ninety-two percent interior monologue. A classroom with a teacher, students, and a boy three rows away she couldn't stop noticing, and nobody spoke for two thousand words.
Dialogue. We measured it: 7.9 percent dialogue across the manuscript, some of the most dramatic chapters under three. Never Let Me Go runs around 18–22 percent; Twilight closer to 25–30. The fix wasn't adding lines; it was rebuilding scenes around an Observation–Interaction–Processing cycle — she observes, someone speaks, she processes what was said. The interiority didn't shrink. It was given voices. The AI proposed the architecture; I wrote the words.
Continuity. This is the part nobody warns you about. Characters at a school table on a Sunday. A running joke whose Tuesday punchline didn't match its Thursday callback. Day counts that felt right — thirty-two, thirty-nine — landing on Saturdays when the scene needed a Monday. Each scene is internally coherent, so the model doesn't catch these on its own; they only surface reading across chapters against a calendar that exists in the world but in no single paragraph. I found most of them. Each fix rippled forward.
Epigraphs. Every chapter opens with a quote from a real work and a comment line beneath it in an unnamed voice — a progression across the book, with the comment lines forming a second conversation that only becomes legible on re-read. Elegant in the abstract. In practice: an opening poem that put eighteen lines between the reader and the protagonist (moved to the back), a deliberately epigraph-free chapter that read as an oversight, a quote that collided with a physical book inside the same chapter. Every change touched three others. Four dedicated sessions.
Mechanical. A style audit where deliberate callbacks had replicated into tics: "without meaning to" cut from eleven to five, each survivor now load-bearing; "the quality of someone who" from twenty-four to twelve; "she had been" from seven per chapter to four. Seven hundred and sixty-eight straight quotation marks that crept in through round-trips between environments. An expansion pass toward the 70,000-word target, where every addition had to do the same kind of work as the prose around it or it didn't belong.
Editorial simulation. The thing the AI is best at, and it deserves its own post. A simulated panel — commercial agent, genre specialist, literary-quality reader, series-engine analyst — each reading the manuscript and returning structured feedback, plus two structural-function passes. They disagreed. The commercial voice wanted more physical sensation in the romantic scenes; the literary voice wanted less. What made it work was that the output was diagnosis, not prescription. The AI told me what wasn't working; I decided how to fix it and wrote the fix. Two or three rounds per chapter before it approached readable.
Metaphor recalibration. The filing system was appearing eight to ten times a chapter; the target was three or four. You can't delete your central metaphor. Each cut had to be replaced by what the metaphor was standing in for — the actual observation, the actual failed response, the moment she reaches for the drawer and the drawer isn't there.
Title change. A full read, not find-and-replace — every instance checked for whether the new title carried the same weight in that sentence.
The pass that's mine
After all of that, the final pass on every chapter is mine alone, and it's often the most extensive. The AI's sentences land correctly and don't sound like a person wrote them. Technically beautiful, and not what this character would say. Right information, wrong music.
I've since come to understand why. The gap isn't that the model can't write physical, specific prose — with enough of my own revised chapters extracted into reference documents, it can. The gap is selection. It can generate twenty plausible details for a moment; it can't reliably pick the one that's alive. A writer sees that one before the sentence starts. So the process became a loop: I revise until the voice is right, the working details from those chapters get extracted and fed back, the AI drafts from that narrower field, I revise again and add what it couldn't find, and those additions go back into the reference. Each cycle raises the model's ceiling and lowers my revision load. The gap narrows. It doesn't close.
What hundreds of passes means
I don't mean I wrote the same chapter hundreds of times. I mean the chapter, and the manuscript around it, received hundreds of different kinds of attention. Structural demolitions. Single-word fixes that took an hour to locate and thirty seconds to make. A whole session debating whether a character's silence was ethically motivated or narratively convenient, which changed three lines of dialogue and nothing else.
What actually happened is something we don't have good language for yet: holding an entire architecture in a shared space, arguing with it, measuring it, breaking and rebuilding it, at a pace and depth impossible alone. It requires a second mind in the room — one that has read everything, remembers everything, and will tell you your favourite sentence is the one that has to go.
It's not automation. It's a collaborative, 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.