How to Write a Novel With AI (Without Losing Your Voice)

The question I get asked is "can AI write my book." The useful question is narrower: which parts of writing a novel can you hand over? And which ones break if you do?

Six months of drafting this way gave me a fairly clear answer. Some of it surprised me.

One thing to flag before anything else. Every guide you will find on this subject is published by a company selling an AI writing tool. That includes this one. The claim you will see repeated across most of them is that AI makes writers finish first drafts up to 40% faster. I have not been able to find a study behind that number anywhere. Treat it as marketing until someone shows the working, mine included.

What to hand over, what to keep

This is the whole method. Everything below is detail.

The pattern: hand over what is laborious, keep what is load-bearing. Where writers get into trouble is handing over decisions rather than labor, usually without noticing.

Step 1: Build the story bible before you draft a word

This is the step people skip and it is the one that decides whether any of it works.

An AI drafting your chapter twelve has no idea what happened in chapter three unless you have given it somewhere to look. Without that, every scene is written by someone who has not read your book.

What actually needs to be in it:

• Premise in one paragraph. What the book is about, not what happens.

• Characters: what each wants, what they are hiding, how they speak differently from each other.

• World rules: anything that must stay true. Magic costs, technology limits, geography.

• Voice: tense, person, register. Two or three sentences of your own prose as a sample.

• What has happened so far, updated as you go.

You will see this called different things depending on what you use. Sudowrite calls it a Story Bible, Novelcrafter a Codex, NovelAI a Lorebook. Ours is Story DNA. Same underlying idea. The name matters far less than whether the thing is actually connected to your drafting.

That connection is the part I would not compromise on. A Google Doc works, as long as whatever you draft with genuinely reads the current version every time. Most people do not, which is why their output drifts around chapter fifteen. Mine lives in Story DNA, so every feature reads it before it writes a word. Because it carries tone and voice rather than only facts, chapter thirty does not quietly turn more formal than chapter three. We have tested it past the 500-page mark without the back half degrading. If you are planning a long book or a series, ask every tool you try for its equivalent number.

The honest version: this takes an afternoon and it is boring. It is also the difference between a tool that helps and a tool that produces plausible nonsense.

Step 2: Outline to the scene, not to the chapter

"Chapter 12: Ada confronts her mother" is not enough to draft from. You will get a competent, generic confrontation.

What works is one line per scene, answering three things: who wants what, what stops them, what is different by the end.

Ada asks her mother about the letter. Mother deflects by talking about the house. Ada leaves without asking again. She now knows her mother is lying.

That is draftable. The first version is not. Almost every complaint about AI writing being bland traces back to a prompt that did not say what the scene was for.

Step 3: Draft scene by scene, never chapter by chapter

Long generations drift. The first 300 words sound like your book, the last 300 sound like a novel-shaped product.

So: one scene at a time, with the story bible in context. I use Draft for this, which takes the outline line and produces the scene rather than the whole chapter.

Then (and this is not optional) rewrite the first paragraph yourself before reading on. It re-establishes your voice in your own head before you start editing. Skip it and you start unconsciously matching the machine's rhythm, which is how people end up with drafts that are technically theirs and do not sound like them.

Step 4: Edit in your voice, not toward it

The difference matters. Editing toward your voice means nudging the generated text until it is close enough. Editing in your voice means treating the generated text as raw material you would not defend.

Practically, on any passage that matters:

• Cut the first sentence. AI scene openings almost always warm up before they start.

• Cut every adverb in dialogue tags. "She said nervously" is the machine telling you what it failed to show.

• Rewrite the last line yourself, always. Scene endings are where voice lives. They are the thing a reader carries into the next scene.

For a phrasing I can feel is wrong but cannot fix, I generate a few alternatives with Rewrite and pick, which is a different activity from accepting. Choosing between options keeps the judgment with me.

Step 5: Check continuity as you go, chapter by chapter

Leave it until the manuscript is finished and you inherit roughly forty small contradictions with no record of which ones you meant.

I ask Talk to your book direct factual questions: what color are Ada's eyes, how many days since the funeral, has she met this character before. Questions about the story, not opinions about the writing.

The two that catch the most errors, every time: how much time has passed, then who knows what right now. Timeline and information state are where long drafts fall apart. Neither is visible when you are reading forward.

The four mistakes I see most

Three of these I made myself.

• Generating a whole chapter at once. It drifts. The first 300 words sound like your book, the last 800 sound like a product. Scene by scene, always.

• Accepting rather than choosing. The moment you take output because it is good enough rather than because you picked it over alternatives, authorship starts leaking. Generating three options and rejecting two keeps the judgment with you.

• Skipping the story bible because you are keen to start. This is the expensive one. Every hour saved here costs several later in continuity repair.

• Letting it write the emotional scenes. Covered below. It is the mistake that shows most in the finished book.

Where AI genuinely does not help

I want to be straight about this, because the tools are not sold this way.

• Deciding what the book is about. If you do not know, AI will happily generate a book about nothing, fluently, for 90,000 words.

• Anything with real emotional weight. Generated grief reads like grief described by someone who has read about grief. Write those yourself. They are the reason anyone finishes the book.

• Knowing whether it is any good. No tool can tell you. Readers can.

• Your first novel, if you want to learn craft. Handing over the hard parts is efficient and you will not learn what those parts taught you.