Guide

How to write a novel with AI

Most AI novels fail in the same three places: the plan, the memory and the prose. Fix those and you have a draft worth editing. Here is the method, step by step, with what the market actually rewards.

Neubook Write · 11 September 2026 · 13 min read

Why most AI novels fall apart

A language model will happily write chapter one. The trouble starts at chapter seven, when the protagonist's sister changes name, the timeline slips by a season, and every scene begins with the weather. This is not a failure of the model's prose; it is a failure of everything around the prose. Three specific failures account for almost all of it.

It has no plan, so it improvises

Asked to "write a novel about a retired postman", a model invents as it goes. Improvisation is fine for a scene and fatal across 80,000 words, because nothing set up in chapter two gets paid off in chapter nineteen. Novels are structural objects. The structure has to exist before the sentences do.

It forgets

Even with a long context window, a model asked to write chapter twelve without being shown what happened in chapters one to eleven will invent a plausible past rather than recall the real one. Continuity is not something the model maintains for you; it is something your process maintains for the model.

It writes in a recognisable register

Readers have learned the patterns. The triads. The "it isn't X, it's Y" reversal. The rhetorical question at the top of a scene. The paragraph that ends on a tidy moral. None of these is wrong in isolation; the density is the tell. Research on AI-assisted writing has found that model help tends to make a group's output more similar to each other even as each individual piece improves, which is exactly the problem for fiction, where difference is the product.

The fix for all three is the same shape: decide the book before you draft it, carry the decisions forward mechanically, and edit the prose against a named list of patterns rather than by vibes.

Decide the book before you draft it

Before any chapter is written, you should be able to answer these in a sentence each. Write them down; they become the brief you hand to the model, and they are what you check the draft against.

• Whose story is it, and what do they want? Not "she wants happiness" — "she wants the tea estate to survive one more harvest so she does not have to sell to the man who left".
• What stands in the way? A person with their own reasons beats an abstract force every time.
• What changes by the end? If nothing internal changes, you have a sequence of events, not a novel.
• What is the promise of the premise? The reason a reader picks it up. In genre fiction this is close to non-negotiable; a romance that does not deliver the couple is a broken contract, not a subversion.
• How long, in chapters? A commercial novel usually lands between 70,000 and 100,000 words. Eighteen to twenty-four chapters at 3,000 to 4,000 words each is a workable shape.

If you cannot answer the third question, stop. That is the book, and no drafting tool will find it for you.

The character sheet that stops the drift

The single highest-leverage artefact in AI-assisted fiction is a short, factual sheet the model sees every time it writes. Not backstory prose — facts that can contradict each other.

Keep it to a page for the whole cast. A sheet nobody reads is a sheet the model will contradict.

Outline at the scene level, not the chapter level

"Chapter 9: they argue" produces nine hundred words of nothing. The outline entry that produces a usable chapter names four things: where it happens, who wants what, what goes wrong, and what is different at the end. Four lines per chapter, twenty chapters, and you have the book's spine in eighty lines.

Two structural notes worth having in front of you while you outline:

• Alternate pressure and release. Readers need the beat after the set piece, or the set pieces stop landing.
• Put the midpoint reversal in deliberately. Around chapter ten or eleven, something the protagonist believed should turn out to be false. A model will not place this for you unless the outline says so.

Draft chapter by chapter, with the story carried forward

Write one chapter per request, and give the model three things each time: the character sheet, a compressed summary of what has happened so far, and the outline entry for this chapter and the next. The last of these matters more than it looks — a chapter written without knowing where the next one goes tends to resolve everything, which is why AI chapters so often feel like short stories bolted together.

Practical settings that change the output

• Ask for scenes, not summary. The default register of a model is summary. "Write the scene in real time, with dialogue" is the most useful sentence in the prompt.
• Name the point of view and the tense, every time. Drift here is common and expensive to fix later.
• Give it something specific of yours. One remembered detail per chapter — a smell, a piece of local knowledge, a thing your aunt says — does more for the prose than any style instruction.

Dialogue that does not sound written

Dialogue is where machine prose is most obvious and most fixable. Models write dialogue in which everyone is articulate, nobody interrupts, and each line answers the one before it. Real speech is worse than that, and better.

Four rules that fix most of it

• Let people answer the wrong question. "Did you call her?" "There's rice in the pan." That exchange carries more than a paragraph of interiority.
• Cut the first line of every exchange. Scenes usually start one beat early. The greeting is almost never the scene.
• One speaker should want something the other will not give. If both characters want the conversation to go well, there is no scene.
• Use "said" and stage business. Models reach for "he articulated", "she countered". Replace with said, or with an action: she put the lid back on the pan.

A practical trick: ask for the scene with a constraint the model cannot smooth over. "Write this argument in under 200 words, no dialogue tags other than said, and neither of them says what they actually mean." Constraints produce better prose than adjectives about style.

The line edit that removes the tells

Drafting produces a manuscript. Editing produces a book. Run every chapter through a pass whose only job is to remove the machine patterns, then read it yourself with a pen. The patterns worth hunting, in rough order of how loudly they announce themselves:

1. The reversal. "It wasn't about the money. It was about being seen." Cut or rewrite; one per book is a style, one per page is a tic.
2. Triads. Three adjectives, three clauses, three examples, endlessly. Vary to one or two.
3. Sentences that start with "You". The model's instinct is to address the reader mid-scene.
4. Em dashes in pairs. Two a chapter is plenty; models write six.
5. The closing moral. A paragraph that stops to explain what it meant. Delete the last sentence and the scene usually improves.
6. "Delve", "tapestry", "testament to", "navigate the complexities". Vocabulary that has become a signature.

You can count these rather than guess at them. Our free AI tells checker scores any passage per thousand words and shows the sentences that trip. And if you are worried about detectors rather than readers, read what the research actually says about AI detectors first — they are unreliable enough that chasing a score is wasted effort, while the patterns above are what a human notices.

Revision in three passes, in this order

Editing everything at once produces a manuscript that is heavily polished and structurally broken. Do one thing per pass.

Pass one: structure, with the book closed

Write a one-line summary of every chapter from memory. The ones you cannot remember are the ones to cut or merge. Check that the midpoint reverses something, that the protagonist causes the ending rather than receiving it, and that every subplot resolves or is deliberately left open.

Pass two: scene by scene, on paper

Read start to finish without fixing sentences. Mark only: where you skimmed, where you were confused, where you stopped believing it. Skimming is a pacing problem, confusion is a clarity problem, disbelief is a character problem. Fix in that order.

Pass three: line edit

Now the sentences. Tells pass first, then your own read for rhythm. Read the first and last line of every chapter in sequence; they should form a spine that makes sense on its own.

If you only have time for one pass, do pass two. Readers forgive a clumsy sentence and abandon a boring chapter.

What the market rewards, if you are writing to sell

Fiction is not one market. If the commercial outcome matters, the genre choice is the single biggest lever you pull, and the data is not subtle. Romance accounts for a large majority of Amazon's most-sold ebooks, and analysis of the Kindle bestseller lists finds that roughly two thirds of best-selling romance ebooks come from indie publishers — the genre where self-publishing has most completely displaced the traditional houses at the commercial top. Subscription reading is a large part of that: most of those titles are enrolled in Kindle Unlimited, where pages-read revenue rewards books readers finish.

Three implications for how you write, not just what:

• Deliver the genre promise exactly. Readers of a category buy the beats. Subvert the execution, not the contract.
• Write to be finished. If revenue comes from pages read, a saggy middle costs money as well as reviews.
• Series beat standalones. The economics of indie fiction are built on a reader finishing book one and buying book two the same night.

The blurb, the title and the first page

Three pieces of text decide whether anyone reads the book you spent a fortnight on, and none of them is the book.

• The blurb is not a summary. It is the promise of the premise plus a complication, in about 120 words, ending on the question the book answers. Write it before you draft; if the blurb is boring, the book will be.
• The title signals genre before it signals content. Look at the top twenty in your category and notice the pattern — the length, the syntax, whether they use a name. Match the pattern, then differentiate inside it.
• The first page is what the Look Inside sample shows. Start in a scene, with a specific person doing a specific thing. No weather, no dream, no waking up.

A blurb skeleton that works

[Character] has [situation] under control.
Then [inciting event].
Now [what they must do], but [cost or obstacle].
[One line that names the stakes as a question.]

Metadata: the part most first novels get wrong

A good book with bad metadata does not get found. Amazon gives you three browse categories and seven keyword fields, and its own help pages are specific about how to use them: choose categories that accurately describe the book and balance popularity against relevance, and use two- to three-word keyword phrases rather than single words, targeting the specific things readers type. Amazon's metadata guidelines also set out what will get a listing suppressed — keyword stuffing in the title, other authors' names in your subtitle, and similar.

A workable keyword set for a novel

• Two that name the trope ("enemies to lovers", "second chance romance")
• Two that name the setting or occupation ("small town romance", "highland historical")
• One that names the heat or tone ("clean and wholesome", "dark romance")
• One that names the format ("standalone novel", "series starter")
• One for the reader, not the book ("books for fans of")

Disclosure, copyright and the honest bits

Publishing a book drafted with AI is allowed on every major platform; being vague about it is what causes trouble. Amazon's content guidelines ask you to declare AI-generated content when you publish, and the declaration is not shown to readers. We wrote a fuller guide on what KDP asks and how to answer it.

On ownership: the US Copyright Office's 2025 report on copyrightability concluded that material generated purely from prompts is not protected, while your selection, arrangement and edits are. The practical reading for novelists is in our guide to copyright on an AI-written book. Put a plain line on the copyright page and move on.

One more, less legal and more professional: if you intend to submit anywhere that forbids AI-assisted work, do not. Magazines have closed submissions over the volume of machine-written stories, as Clarkesworld did, and misrepresenting the work burns a relationship you cannot rebuild.

What it costs to do this properly

Drafting with AI collapses the cost of the first draft to near zero and moves the expense to the parts that were always the expensive parts. A realistic budget for one indie novel you intend to sell:

The pattern that separates the authors who earn from the ones who do not is rarely the drafting method. It is that they spend on the cover and the edit, and they publish the next book.

A two-week schedule that actually finishes

The only step people skip is day 4 to 9, and it is the one that decides whether the book is any good.

Sources

• KDP: Categories — how Amazon shelves your book
• KDP: Make your book more discoverable with keywords
• KDP: Metadata guidelines for books
• KDP: Content guidelines, including AI-generated content
• US Copyright Office: Copyright and Artificial Intelligence, Part 2 (2025)
• Romancing the Data: Romance best-seller trends, publishers, price and POV
• Ebook statistics: market, sales and authors
• Science Advances: generative AI assistance and the homogenisation of output
• BuzzFeed News: Clarkesworld closes submissions over AI-written stories

Common questions