In this guide
Romance, thriller, fantasy and horror all share one problem in the middle of the book: the first act made promises that nothing is keeping track of. A missing sister, a debt, a rule about the magic, a character who said they would never come back. By chapter twenty, a human writer has forgotten two of them and a model has forgotten all of them.
That is the real failure mode. Not flat sentences. Lost promises.
The chapter-twenty problem
Ask a model for chapter one and you will get something competent. Ask for chapter twenty with only a vague memory of chapter one and you get a chapter that is competent and slightly wrong: a character's voice has drifted, a rule has softened, a subplot has quietly vanished.
The cause is structural, not stylistic. A conversation compresses the past as it goes. Each summary keeps the plot and loses the texture, and the texture is what the promises were made of.
There are two ways out. The writer can hold the entire book in their head and re-paste what matters, or the tool can store what matters and inject it. The first works until it does not. The second is what a project is for.
What a novel has to remember
Five things, and only one of them is plot:
- The cast. Who they are, what they want, what they fear, how they speak, what they would never do.
- The rules. Magic costs something. The city has no phones. The narrator cannot leave the house. Whatever the book establishes, it has to obey.
- The promises. Setup and payoff: the object introduced in chapter two, the debt, the unanswered question.
- The timeline. Who knows what, and when they learned it. This is where AI novels break most visibly.
- The voice. Tense, person, sentence length, what the narrator notices and what they refuse to describe.
A chapter prompt that carries all five is large. A chapter prompt that carries none is the reason the draft drifts.

Put the cast in facts, not in the prompt
The most useful move is also the most boring: write the cast down as facts before you draft.
In the book's Global knowledge panel, facts are label-and-value pairs that get injected into every chapter. For fiction, useful labels look like this:
| Label | Value |
|---|---|
| Mara wants | To get her brother's name off the lease before the audit |
| Mara fears | Being seen as the person who ran |
| Mara's speech | Short sentences. Never asks a question she does not already know the answer to |
| The rule | Nobody can lie inside the boarding house, but they can refuse to answer |
| Timeline | The audit is nine days away. Chapter one is day one |
Three characters with five facts each is fifteen lines. It will do more for continuity than any instruction to "keep the characters consistent", because it gives the model something to check against instead of an adjective to obey.
Plan escalation, not chapters
A chapter plan that reads like a list of events produces a book that moves without escalating. The type brief for fiction asks for the opposite: each chapter should leave the protagonist further from the goal than they started.
Write the one-line goal of each chapter as a change, not a topic:
- Weak: "Mara talks to the landlord."
- Better: "Mara trades the deposit for a week, and learns the landlord already knows about the audit."
The second version gives the model a starting state and an ending state. It also gives you a way to check the plan before you write it: read the goals in order and ask whether the pressure rises. If chapter twelve has the same stakes as chapter four, the middle will sag, and no prose style will save it.
Plans run from 3 to 60 units, and regenerating appends to the existing plan. When the middle sags, rewrite the goals from chapter ten onward and regenerate those units. The first nine chapters survive.
Use the recap as a continuity check
After each chapter, the editor writes a short recap — what happened, what changed, which concepts were established — and the next chapter is drafted against it instead of against the whole manuscript. That is what keeps the context small enough to be affordable and focused enough to be useful.
Your job is to read those recaps once a week. They are a compressed version of your own book, and mistakes show up there before they compound:
- A recap that says a character learned something they already knew in chapter six.
- A recap that promotes a minor character to a decision-maker.
- A recap that resolves a promise you meant to keep open.
Fixing the recap costs a sentence. Fixing the chapter costs a rewrite.
Where to use Continue, and where to stop
Continue extends the scene in front of you. It is the right tool when a chapter stopped one beat early: the confrontation has started but not landed, the reveal is set up but not delivered. Give it a steering note — "end on the sound of the lift arriving" — and it will finish the scene rather than starting a new one.
It is the wrong tool when the scene has already resolved. Continuing a finished scene produces a coda, then another coda, and the chapter stops moving. If the model has already delivered the change the chapter goal asked for, stop and start the next chapter.
The prose problem no setting fixes
Everything above keeps a novel coherent. None of it makes the prose good, and fiction is where readers are least forgiving. Our own publishing note for the fiction type says it plainly: fiction lives or dies on prose quality, AI disclosure hits hardest here, and readers judge the first page only.
That is not a reason to avoid the tool. It is a reason to use it honestly:
- Draft with the model, then rewrite the first page by hand until it sounds like you.
- Keep one scene per chapter that only you could have written: a place you know, a job you have done, a specific kind of silence.
- Use the model to interrogate, not to decide. Ask it what a reader would expect to happen next, then do something else on purpose.
A workflow that survives forty chapters
- Premise, in one paragraph. Protagonist, want, obstacle, what winning costs. If you cannot name the cost, the book has no engine.
- Facts. Five lines per principal character, three rules of the world, the timeline's zero point.
- Plan, 40 to 60 units. One-line goals written as changes. Read them in order for escalation before drafting anything.
- Draft three chapters, then stop. Read them together. Does the voice hold? Does the timeline hold? Fix the facts before continuing.
- Draft in blocks of five. Read the recaps after each block. Rewrite plan goals when the pressure drops.
- Rewrite the first chapter last. It is the page readers judge, and by then you know what the book is actually about.
At 1,200 words a chapter, 70,000 words is about 58 chapters. Plus covers roughly 660 chapters a month, so the drafting is not the constraint. The constraint is the reading and the rewriting, which is exactly where a novelist's time should go.
The bottom line
An AI novel writer is useful when it holds the parts of a novel that are expensive to lose: the cast, the rules, the promises and the timeline. Give it a spine made of facts and escalating chapter goals, use the recap to catch drift, and keep the first page and the best scene for yourself. The model can write a chapter. The novel is still the part only you can decide.
Key takeaways
- 1Novels fail in the middle because promises are lost, not because sentences are flat.
- 2Store the cast, the world’s rules and the timeline as facts the model can contradict.
- 3Write chapter goals as changes so escalation is visible in the plan.
- 4Read the recaps weekly: fixing a recap costs a sentence, fixing a chapter costs a rewrite.
- 5Use Continue to finish a scene, never to extend one that has already resolved.
Frequently asked questions
Can an AI write a whole novel?
It can draft one, at roughly 58 chapters for 70,000 words at 1,200 words a chapter. Whether the result is worth reading depends on the prose and the choices, which are still yours. Fiction is the hardest category to publish with AI, and readers judge the first page only.
Is AI-written fiction detectable?
Often, yes, especially in the first pages. No tool can promise otherwise. The better use of the technology is drafting and revision support while you keep the voice, the specifics and the judgment.
How many chapters should a novel have?
In our editor, a 70,000-word fiction book is about 58 chapters at the 1,200-word default. You can set chapters between 400 and 4,000 words, so the number is a planning choice, not a rule.
Sources
- Google Search Central — Creating helpful, reliable, people-first content — read on 2026-11-16
About the author
One useful essay, every other week
Notes on writing, publishing, and the unglamorous parts of shipping software. No roundups, no growth hacks.




