How to Write a Novel With AI in 2026: A Working Writer's Workflow

This is the workflow I actually use and teach: five stages, each with a clear job for you and a clear job for the AI. It works in any tool that lets you keep a story bible and feed it into the model, whether that is SparkLore, Sudowrite, Novelcrafter, or a folder of text files and a chat window.

The honest starting point

Before the workflow, one fact worth internalising. Current models are very good at local tasks: a paragraph, a scene, a line of dialogue, a list of options. They are weak at global tasks: holding 90,000 words of causality in mind, remembering that your protagonist lost her left hand in chapter four, keeping a voice steady across months of sessions.

Users of every major tool say the same thing. A Capterra reviewer of Sudowrite wrote that it "loses track of the characters, storyline and past events," and a Reddit line quoted by BuiltWritten complains that "the AI sometimes forgets my protagonist has blue eyes, even when it's in the Bible." That is not a bug in one product. It is the shape of the technology, and the workflow below is built around it: you own the global, the AI helps with the local.

Stage 1: Premise (one afternoon)

Your job here is to decide what the book is. Not the plot, the book: who wants what, what stands in the way, why it matters, and what kind of reading experience you are promising.

Write a premise in this form and refuse to move on until every slot is filled:

A [protagonist with one defining trait] must [concrete goal] before [deadline or pressure], but [obstacle], and if she fails, [stakes that cost her something she cannot replace].

A weak version: "A young woman discovers she has magic and must save her kingdom." A working version: "A court archivist who has spent ten years erasing her family's name must forge a royal succession document before the coronation, but the only witness who can authenticate it is the brother she betrayed, and if she fails the archive burns with everything she is."

What AI is good for at this stage: generating thirty premises so you can find out what you do not want. Ask for variations on tone, on protagonist profession, on the shape of the stakes. Read them fast. The one that annoys you because it is close but wrong is often the seed.

What AI is bad for: choosing. A model will happily tell you a premise is "compelling and high-stakes." It says that about all of them.

Stage 2: Story bible (two to five days)

This is the stage most people skip, and it is the one that determines whether AI drafting helps or hurts. A story bible is the file that lets a model behave as if it remembers your book.

Mine has six sections, and I have written up the full template separately:

Write it in plain prose, not bullet fragments. Models read prose better, and so will you in month four.

The bible is a living document. Every time the draft reveals a fact ("Mara is allergic to horses, apparently"), it goes in the bible the same day. This discipline is what separates writers who get consistent output from writers who fight the tool.

Stage 3: Outline (one week)

There are five outlining methods worth knowing, and I compare them in How to Outline a Novel. For AI-assisted drafting, the outline needs one property above all others: each entry must be a scene with a change in it.

Not this:

Chapter 7: Mara travels to the capital and meets the brother.

This:

Ch 7, Mara POV, the river road, dusk. Goal: reach the capital unseen. Obstacle: a checkpoint run by her brother's men. Turn: she is recognised not by a guard but by her brother's dog. Ends with her under escort, not under arrest, which is worse. New fact: the brother knows she is alive.

That second version is something a model can draft from, and something you can drop into a prompt with the bible and get a usable scene. The first version gets you 1,200 words of travelogue.

Use AI here for gap-finding. Paste the outline and ask: "Which scenes could be cut without breaking causality? Where does the protagonist stop making choices? Which secondary character disappears for more than 30 pages?" It is a competent, tireless structural reader. Use it as one; do not let it rewrite your outline to a template.

Stage 4: Drafting (the long middle)

Here is where workflows diverge, and where you should be honest about how you like to write. I see three patterns work.

Scaffold and rewrite. You give the model the bible slice plus the scene card and ask for a draft. You then rewrite every sentence. The AI output is a sketch: it settles blocking, pacing and beats so you are never staring at a blank page. Fast, but only if you genuinely rewrite. Writers who "lightly edit" produce books that read like everyone else's.

Draft and expand. You write the scene in compressed form, dialogue and action only, maybe 400 words. You ask the model to expand description and interiority in your voice, using the style samples in the bible. Then you cut most of what it added. Good for writers who hear dialogue first.

Draft and critique. You write everything. After each scene you ask the model three specific questions: What does the reader know at the end that they did not at the start? Where does the tension drop? What did this character want and did they get it? No prose from the AI at all. Slowest, cleanest.

Whichever you use, keep these rules:

1. Feed the context every time. The relevant characters, the setting rule in play, the previous scene's last 300 words, and the scene card. Tools with a built-in bible do this automatically; in a chat window, you paste it.
2. Ask for less than you want. Request 500 words and extend; do not request 3,000. Long generations drift.
3. Never accept a new fact from the model. If it says the brother has a scar, decide whether that is true. If yes, into the bible. If no, delete it now, not in revision.
4. End the session with a summary. Two sentences on what changed, appended to the timeline. Tomorrow's you will thank you.

A note on pace: a scaffold-and-rewrite writer can reasonably hold 1,500 to 2,500 finished words a day. That is not ten times faster than drafting alone. It is maybe twice, and the second half of the gain comes from never being stuck, not from typing less.

Stage 5: Revision (as long as it takes)

AI is more useful in revision than in drafting, and almost nobody uses it that way.

Continuity pass. Paste each chapter with the bible and ask for contradictions. Eye colour, weather, who knows what, the day of the week. Models catch these well because it is a local matching task.

Tension map. For each scene, ask for a one-line summary and a 1 to 5 tension score. Put the scores in a spreadsheet. Flat stretches show up instantly.

Line-level tells. If you used AI in drafting, you now need to remove its fingerprints: the even paragraph lengths, the "a testament to," the character who "felt a surge of." I keep a list in How to Edit AI-Generated Prose. Search-and-destroy is faster than reading for them.

Voice check. Take a paragraph you are proud of from chapter one and one from chapter twenty. Read them aloud. If they sound like two writers, the AI won in the middle and you need to go back through.

Then the 30-point checklist before anyone else reads it.

A before-and-after from the middle of the process

Scene card: Mara, escorted into her brother's house. Goal: find out what he wants. Turn: he wants nothing; he is dying.

Raw scaffold from a model, fed with the bible:

The house was exactly as she remembered it, and that was somehow worse. Her brother sat by the window, thinner than she had ever seen him. "You came back," he said, his voice heavy with emotion. Mara felt a wave of guilt wash over her. She had betrayed him, and now here he was, dying.

Serviceable. Everything is there and nothing is alive. After rewriting:

Nothing in the house had moved in ten years, which meant someone had been paid to keep it from moving. Her brother was in the window chair. He had always been the big one. "You got thin," she said, because it was the only true thing she could say first. He laughed, then coughed, then held up a hand so she would not come closer, and she understood the escort had not been for her protection.

The scaffold told me where the scene turned. The rewrite decided what it was about: that she reads his illness before he says it, that the escort was to keep her from him and not him from her. The model did not know that. It could not have.

What to watch out for

• Credit surprises. Several tools bill by opaque credits; Sudowrite's own docs admit the cost per action is unpredictable, and it draws some of that product's harshest reviews (research summary). Know how you are billed before a long drafting month.
• Tool lock-in. Keep the bible and manuscript in a format you can export. Plain text and Markdown outlive every app.
• Genre filters. Some tools restrict content unpredictably after purchase. If you write dark material, test before you pay.
• Your own drift. The biggest risk is not the model forgetting your character. It is you, in month three, accepting a scene you would have rejected in month one because it is there and it is fine. Fine is the enemy.

Try it in SparkLore

SparkLore is built around this workflow: a project-level story bible that is injected into every generation, scene cards you draft from, and a continuation tool that reads your last pages before it writes. If you want to test the drafting stage before signing up for anything, the free AI story continuation tool takes a passage and a few bible notes and writes the next beat, so you can see for yourself how much difference the context makes. When you are ready to build the whole bible, create a free account and start with the premise.