AI for Book Writing: My Complete Process

Someone recently told me my books read like they were written by a person, not a machine. They meant it as a compliment, and I took it as one. But the subtext was clear: they assumed AI-assisted meant AI-generated. It does not. The distinction matters enormously, and it determines whether you produce something worth reading or something that adds to the growing pile of forgettable AI slop on Amazon.

I have written six books with AI assistance. Four business books in the “Subtract to Ship” series and two magic performance books in “Late to the Table.” Every one of them contains my ideas, my experiences, my voice, and my editorial judgment. AI handled the parts of book production that do not require those things. This separation is the entire process, and getting it right is what makes the books work.

The Distinction That Defines Everything

An AI-generated book is one where the author provides a topic and the AI produces the content. The author is essentially a curator, selecting and organizing AI output. These books are easy to spot: they are competent, generic, and say nothing the reader could not find in a Wikipedia article or a blog post roundup.

An AI-assisted book is one where the author provides the expertise, ideas, stories, and editorial judgment, and the AI handles the mechanical production work. The author is the thinker. The AI is the typist on steroids.

The practical difference shows up in specificity. AI-generated books deal in generalities because AI only knows generalities. AI-assisted books are full of specific experiences, original frameworks, and opinions that come from one person’s unique perspective. When I write about building a business in Austria, the specifics come from my twenty years in the Austrian business environment. The AI could not generate those specifics because it has not lived them.

My rule of thumb: if a section could have been written by any knowledgeable person in the field, the AI can draft it. If a section requires my specific experience, perspective, or judgment, I write it myself. About thirty-five percent of my books fall into the second category, and that thirty-five percent is what makes them mine.

Before you start your book project, ask yourself: what do I know, think, or believe that is genuinely mine? That is your book. Everything else is production.

Phase 1: The Source Material Audit

A book starts long before the first word is written. It starts with what you already know.

I spent the first two weeks of my book project compiling source material. Not researching from scratch. Gathering what I already had: twenty years of consulting notes, workshop materials, Startup Burgenland documentation, client case studies, magic performance journals, and rough frameworks I had developed over the years.

The total was roughly 3,200 source pieces. Most people have more raw material than they realize. If you have been working in your field for more than five years, you have a book’s worth of material scattered across emails, presentations, notes, and your memory.

AI’s role in this phase: organization. I fed batches of source material into Claude using a specific structure that makes a real difference with long documents:

<source_material>
  [All source documents placed at TOP of the prompt]
</source_material>
<instructions>
  First, quote the 5-10 most significant passages from the source material 
  that represent core ideas or frameworks.
  Then cluster related ideas into thematic groups.
  Identify themes, surface contradictions, and flag gaps in the thinking.
  Note which clusters have the richest source material and which are thin.
</instructions>

Two techniques matter here. First, placing source material at the top of the prompt with instructions at the bottom. With book-length source material, this produces up to thirty percent better results than the reverse. Why? The AI reads the material, builds an internal model of it, and then encounters the task with full context loaded. Put instructions first and the AI starts working before it fully understands the material.

Second, the “quote before analyzing” instruction. By requiring the AI to quote specific passages before clustering, it stays anchored to your actual material rather than generating generic thematic groupings. This is especially important when you have 3,200 source pieces — without anchoring, the AI defaults to broad categories that could apply to any business book.

The AI did not add knowledge. It organized existing knowledge into a structure I could see.

My role: deciding which material was genuinely valuable and which was noise. The AI flagged everything. I decided what mattered. The subtraction audit applied directly here: I cut about forty percent of the material that did not serve the core argument of each book.

Your action: before writing anything, spend one week gathering every piece of relevant material you have. Upload it to AI in organized batches using the structure above. The patterns that emerge will show you your book’s structure more clearly than any outline exercise.

Phase 2: Architecture, Not Outlining

I do not call this phase outlining because outlining implies a linear process. Book architecture is structural. It is about how ideas relate to each other, where the reader’s understanding builds, and what sequence creates the clearest possible argument.

For each book, I created what I call an argument map. Not a list of chapters, but a visual representation of how the core ideas connect. Each node is a key insight. The connections show dependencies: which ideas must come before others, which ideas reinforce each other, and which ideas are independent.

AI helped me test the architecture. I would describe the planned structure and ask: “If a reader encounters idea X in chapter 3, will they have enough context from chapters 1 and 2 to understand it? What background knowledge am I assuming?” The AI identifies assumption gaps that are invisible to the expert author because you forget what you know that your reader does not.

I also used AI to stress-test chapter sequences. “Here are the key points of chapters 4 through 6. Is there a more logical order? Where might a reader get confused or lose the thread?” The AI suggested reordering in three of the four business books, and in two of those cases the suggestion was clearly right.

Your architecture phase should take about a week per book. Resist the urge to start writing before the architecture is solid. A well-structured book with average prose is better than a poorly structured book with brilliant prose. Architecture determines whether the reader follows your argument. Prose determines whether they enjoy following it.

Phase 3: The Writing Process

This is where AI contributes the most volume and where the process requires the most discipline.

For each chapter, my writing process follows a consistent pattern:

Step 1: Chapter brief. I write a one-page brief for the chapter: thesis statement, key supporting points (three to five), planned examples and stories, target word count, and how this chapter connects to the ones before and after it. This brief is my authorial intent. The AI receives it as direction.

Step 2: Section-by-section drafting. I do not prompt “write chapter 7.” I work through each section individually using XML-structured prompts with few-shot examples:

<section_brief>
  Topic: Pricing psychology -- the Vulpine experiment
  Target: 500 words
  Tone: Conversational, slightly self-deprecating, building to the insight 
  that raising prices increased conversion
</section_brief>
<source_material>
  [The specific Vulpine pricing data and my notes on what happened]
</source_material>
<few_shot_examples>
  <example>
    Last November I raised Vulpine's base package from EUR 2,400 to EUR 3,800. 
    Three clients left. Seven new ones signed within six weeks. The math was 
    not subtle.
  </example>
  <example>
    I had been looking at the ground floor of a skyscraper and concluding 
    that I understood the entire structure.
  </example>
</few_shot_examples>
<instructions>
  Write the opening section. Match the voice and rhythm of the examples.
  Quote relevant details from the source material before generating prose.
</instructions>

Few-shot examples are the single most reliable steering mechanism for book writing. Two to three examples per section, showing the AI the exact rhythm and specificity level I want, produces dramatically better first drafts than detailed prose instructions alone. The examples communicate patterns that rules cannot capture — sentence length variation, the ratio of observation to analysis, when to be blunt versus reflective.

The specificity matters. The more direction I provide, the closer the draft comes to what I want. Vague prompts produce vague drafts.

Step 3: Personal sections. Some sections I write entirely myself. Stories from my life, controversial opinions, sections where my specific voice matters most. These sections are the heartbeat of the book. The AI cannot write them because they require things only I have.

Step 4: Assembly and flow. Once all sections are drafted, I assemble them and read through for flow. The transitions between AI-drafted and self-written sections need attention. AI tends to write in a slightly different rhythm than I do, and the seams are visible if I do not smooth them.

This process produced roughly 3,000 to 5,000 words per day of first-draft content, which is three to five times faster than I could write from scratch. But I want to be clear: these were first drafts. They were not publishable. The editing phase is where they became books.

Phase 4: Editing as the Real Writing

I believe editing is where the actual book happens. The first draft gets words on the page. Editing turns those words into something worth reading.

My editing process has three passes:

Pass 1: Structural edit. Does each chapter deliver on its promise? Are the arguments logical? Is anything missing? Is anything redundant? This pass often results in moving sections between chapters, cutting entire sections, or adding new material. AI helps by reviewing the full draft for structural coherence, but I make all structural decisions.

Pass 2: Voice and clarity edit. This is where I make the book sound like me. I read every sentence and ask: would I say this? Is this clear? Is this interesting? Sections drafted by AI get the heaviest editing here because even well-prompted AI has a recognizable flavor that needs to be replaced with my actual voice. I typically rewrite twenty to thirty percent of AI-drafted sections during this pass.

Pass 3: Detail edit. Fact-checking, consistency checks, formatting, and polish. AI is genuinely useful here. I use a separate editing agent with structured evaluation criteria:

<evaluation_criteria>
  <criterion name="factual_accuracy">Are all claims verifiable? Flag specific 
  statistics and named studies.</criterion>
  <criterion name="terminology_consistency">Are key terms used the same way 
  across all chapters?</criterion>
  <criterion name="voice_match">Does this chapter sound like the same author 
  as every other chapter?</criterion>
  <criterion name="cross_reference_accuracy">Do all references to other 
  chapters and books point to correct content?</criterion>
</evaluation_criteria>

The agent reviews against these criteria and flags issues. I fix them. Structured evaluation criteria produce more consistent reviews than open-ended “check this chapter for errors” prompts, because the AI checks each criterion independently rather than doing a general impression pass.

The editing phase took longer than the drafting phase. For my six books, drafting took roughly six to seven weeks. Editing took eight to nine weeks. This ratio is correct. If you are spending more time drafting than editing, you are either an exceptionally clean first-draft writer or you are not editing enough.

I have written about the full timeline, but the key point here is that AI did not compress the editing. It compressed the drafting, which freed me to spend more time on the editing that makes a book good.

Phase 5: Production and Launch

The final phase covers everything from completed manuscript to published book: formatting, cover design, metadata, distribution setup, marketing materials, and launch strategy.

AI’s role in this phase:

• Generating back-cover copy variations (I wrote five versions and picked the best)
• Creating chapter summaries for marketing use
• Writing email sequences for the launch campaign
• Producing social media content for pre-launch, launch, and post-launch periods
• Drafting the book description for Amazon and other retailers
• Creating media kit materials

Human decisions in this phase: cover design direction (I hired a designer for final covers), pricing strategy, distribution channel selection, launch timing, and marketing budget allocation.

The production phase benefits from AI because it involves a large volume of written materials that all need to be consistent with each other and with the book itself. Having the same AI that helped produce the book also produce the marketing materials ensures consistency of voice and messaging.

For self-publishing authors considering the Austrian market, the production phase is where AI most levels the playing field with traditional publishers. A traditional publisher has teams for copywriting, marketing, and production. AI gives you equivalent capability at subscription cost.

What This Process Costs

Let me be transparent about the full cost of producing six books with AI assistance:

• AI subscriptions and API costs: approximately EUR 800 over the production period
• Cover design (human designer): EUR 1,200 for six covers
• Editing software and tools: EUR 100
• My time: roughly 500 hours over the production period

The monetary cost was approximately EUR 2,100 for six books. Compare that to traditional ghostwriting (EUR 15,000-30,000 per book) or traditional publishing support (thousands in editing, design, and production per book).

The time cost was significant but concentrated. Five hundred hours is roughly three months of full-time work, spread out alongside other business commitments.

The real cost comparison is not against hiring a ghostwriter. It is against writing without AI. Without AI assistance, these six books would have taken an estimated eighteen to twenty-four months at the same daily time commitment. AI compressed the timeline by roughly seventy-five percent.

Common Mistakes in AI Book Writing

Having talked to dozens of authors using AI, I see the same mistakes repeatedly.

Mistake: Using AI without having expertise. If you do not know your subject deeply, AI will not save you. It will produce a generic book that adds nothing to the conversation. Have something to say before you use AI to say it.

Mistake: Skipping the editing phase. AI drafts are not publishable drafts. Authors who publish lightly edited AI output produce books that readers correctly identify as AI-generated, and they do not come back for book two.

Mistake: Letting AI structure the argument. The structure is the argument. If AI decides the chapter order, the section flow, and the emphasis, it is the AI’s book with your name on it. Author the structure. Let AI help with the prose.

Mistake: Writing the whole book in one voice. Your best writing has variety. Some sections are punchy and direct. Others are reflective and nuanced. AI defaults to one register unless you explicitly direct variation. Build tonal shifts into your chapter briefs.

Mistake: Not quality-checking AI contributions. AI confidently makes factual errors, invents statistics, and attributes quotes to the wrong people. Every factual claim in an AI draft needs verification. This is tedious but non-negotiable if you value your credibility.

Takeaways

1. AI-assisted means you provide the expertise and judgment; AI provides production speed. If you cannot clearly identify what only you can contribute, you do not yet have a book.

2. Spend two weeks gathering and organizing source material before writing a word. Your existing knowledge is your book’s raw material. AI organizes it; you decide what matters.

3. Author the architecture yourself. The structure is the argument. Let AI help test and refine it, but the structural decisions must be yours.

4. Plan for editing to take longer than drafting. AI compresses drafting time dramatically. Editing time stays roughly the same because it requires your judgment and voice.

5. Verify every factual claim in AI-drafted sections. AI confidently generates plausible-sounding errors. One wrong statistic in a published book damages your credibility permanently.