Guide

We wrote a romance novel with AI, and here is everything it took

Not a demo. One complete book, with the brief we gave it, the numbers it produced and the three things that went wrong.

Neubook Write · 4 October 2026 · 4 min read

What we set out to test

Whether a planning-first pipeline could produce a romance novel that a genre reader would accept, and what it would actually cost. One book, written end to end, with the numbers published whether they flattered us or not.

The brief

The interview is where the book is decided. Ours was deliberately specific, because generic answers produce generic books:

• Premise. A woman returns to sell her late father's restaurant; the volunteer fire captain who talked him out of selling it ten years ago is still there.
• The unfinished argument. The sale would have paid for her place at a culinary school in Lyon. He has never known that.
• Texture. Twelve tables, the same octopus with potatoes since 1979, recipes in an exercise book written in handfuls rather than grams.
• Sensory anchors. Woodsmoke that arrives three days before anyone sees flame; the harbour bell that only rings when the wind turns onshore.
• A planted question. His hands are scarred. She notices early and does not ask.
• Directions. Alternating point of view, open door, every chapter ends on a hook, no epilogue that explains the book back to the reader.

What the plan did with it

Every specific came back in the structure, and the plan invented one better than ours: the father's will leaves the restaurant to her alone with a note saying settle it your way. The chapter list placed the planted details deliberately, with the scarred hands noticed at six, asked about at ten, and the consequence landing at fourteen.

The numbers

What the prose looks like

She parks her rental Civic crooked across two spaces because there is no one to tell her not to, and sits for a second with the engine ticking.

And the line the book is built around, in chapter fourteen:

"I didn't save the restaurant for the town," Cal said. "I'm not that good a man."

Which is corrected, rather than repeated, five chapters later: "I said I wasn't that good a man. There's a difference." That structure came from the plan, not from any individual chapter.

Three things that went wrong

1. It came out short. 34,328 words is a short novel bordering on a novella, against a 70,000 to 90,000 genre expectation. The chapters ran 18% under target and nothing caught it until the end. Length is a planning decision and we made it badly.
2. The cover contradicted the book. We made the cover before the book existed, with a placeholder town name, and the finished novel used a different one. Nobody noticed until both were on screen together.
3. Quality checking does not travel. Our machine-prose scoring is English-only, so any translated edition ships with no equivalent measure. Worth knowing before you publish in five languages.

What we would tell someone starting

• Spend your effort on the interview. The specifics you put in are what the book is made of, and no model invents them for you.
• Decide the length arithmetically before drafting: chapters times words, checked against the genre target.
• Plant something early that you refuse to answer. It is the single most effective structural device in the genre and it is easy to specify up front.
• Read the whole book before publishing it. We are saying this because it is the step most easily skipped when a draft arrives finished.

Sources

• Burn Season, the romance novel used for every measurement on this page
• KDP content guidelines on AI-generated content
• Google: creating helpful, people-first content

Common questions