Writing Books With AI: The Twelve-Step Author Stack

The internet is full of people claiming they wrote a book in a weekend. Most of those books are unreadable, and readers can tell within a page. The useful question is narrower and much more valuable: where in the writing process does AI actually earn its place?

Twelve places, in order.

Before you write

1. Market research. Analyse the top books in your category — what promises do the blurbs make, what do three-star reviews complain about? Three-star reviews are the single richest source of unmet reader demand in publishing.
2. Positioning. Draft twenty one-sentence pitches, then cut to the one that makes you slightly nervous. AI is very good at generating options and very bad at choosing.
3. Outline pressure-testing. Give a model your outline and ask what a reader would predict, where the middle sags, and what is promised but never delivered. Take the criticism, ignore the rewrites.
4. Research compression. Feed real sources into a grounded notebook and interrogate them. Never let a model supply a fact you cannot trace — hallucinated citations have ended careers.

While you write

1. Beating the blank page. Ask for three terrible openings to the scene. Reacting to bad work is far easier than starting from nothing, and you will write your own line by the second one.
2. Continuity checking. Eye colour, timeline, who knows what and when. A model with your manuscript in context catches contradictions a human editor charges for.
3. Dialogue variation. Ask how a specific character would refuse an offer, given their history. Use it as a prompt for your own ear, not as final text.
4. Description unsticking. When you cannot see the room, generate five sensory details and keep the one that surprises you.
5. Line editing. Genuinely strong now — repetition, filter words, weak verbs, sentence rhythm. Accept about a third of suggestions and you will improve.

After you write

1. Blurb and back cover. Twenty variants, test on real readers. The blurb sells more copies than the book does.
2. Formatting and production. Atticus or Vellum turn a manuscript into print-ready files in an afternoon.
3. Launch assets. Metadata and keywords for KDP, audiobook narration, a reader magnet, thirty social posts derived from actual passages.

The four places it will wreck your book

• Voice. Generated prose regresses to the mean of everything ever written. Your voice is the only thing a reader cannot get elsewhere. Write your own sentences.
• Emotional truth. Models write what grief looks like from the outside. Readers can feel the difference and they will not forgive it.
• Structure at scale. A model holding a whole novel loses causal threads. You are the architect; it is a very fast bricklayer.
• Facts. Anything checkable must be checked by you against a real source. Every time.

Use AI on everything around the words. Be much more careful with the words themselves.

The stack

Novelcrafter for long-form structure with a story codex. Sudowrite for fiction-specific drafting help. Obsidian for research you own locally. Grammarly for the final pass. ElevenLabs for audiobook production. Draft2Digital for wide distribution beyond Amazon.

One more thing: disclosure rules differ by platform and are changing. Check your retailer\’s current policy before you upload, and keep records of your process. The full breakdown lives on the writing and books page, and the marketing page covers selling it once it exists.