People hear “six books using AI” and assume I am either lying or the books are garbage. Fair enough. Let me address both concerns upfront: the books are real, they are published, and readers consistently rate them 4.5 stars or higher. And no, I did not just press a button and have AI spit out six books. What I did was build a production system that let me focus my time on the parts that require human judgment while AI handled the parts that require scale and consistency.
This is the full breakdown. Not theory. Not principles. The exact process, tools, and workflow I used.
The Six Books and Why They Exist
The books span two series. Four business books under the “Subtract to Ship” methodology, covering how to build and grow businesses by doing less, not more. And two magic performance books under “Late to the Table,” covering close-up magic from the perspective of someone who came to performing later in life.
Why six at once? Because the content was interconnected. The four business books reference each other. The two magic books are companion volumes. Writing them sequentially would have meant constant back-and-forth to maintain consistency. Writing them in parallel, moving between books as ideas connected, produced a more coherent set.
This is not a strategy I would recommend for everyone. It worked because I had twenty years of material in my head waiting to be organized, and because the AI content pipeline I built could handle the throughput.
The Production System
Here is the architecture. Five stages, each with specific AI assistance and specific human oversight.
Stage 1: Research and organization. I started with roughly 3,200 source pieces: notes from consulting projects, workshop materials, Startup Burgenland documentation, magic performance notes, audience feedback, and industry research I had collected over two decades.
AI’s role: I fed these sources into Claude in organized batches and asked it to identify themes, cluster related ideas, and surface contradictions. The AI was not generating content here. It was organizing twenty years of my thinking into a structure I could see clearly.
My role: reviewing the clusters, deciding which themes deserved full chapters, and making the editorial decisions about what to include and what to cut. This is where the subtraction audit applied to the books themselves. I cut roughly forty percent of the material that did not serve the core argument of each book.
Stage 2: Outlining. For each book, I created a detailed chapter-by-chapter outline. Each chapter had a thesis, three to five supporting points, planned examples, and a clear reader takeaway.
AI’s role: I drafted rough outlines and asked AI to identify gaps in logic, suggest where additional examples would strengthen arguments, and flag chapters that overlapped too much.
My role: making all structural decisions. The AI suggested. I decided. This distinction matters because the structure of a book is the argument, and the argument has to be mine.
Stage 3: First drafts. This is where AI contributed the most volume and where the process gets the most skepticism.
For each chapter, I provided the AI with XML-structured context that separated concerns clearly:
<source_material>
[Chapter outline + relevant source materials placed at TOP]
</source_material>
<voice_guide>
[Book's voice parameters -- tone, rhythm, vocabulary]
</voice_guide>
<few_shot_examples>
<example type="opening_scene">
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.
</example>
<example type="analytical">
Your brain forms beliefs by averaging your environment. If zero people
around you have started a business, your brain literally cannot model
it as achievable. This isn't weakness. It's biology.
</example>
<example type="reflective">
I had been looking at the ground floor of a skyscraper and concluding
that I understood the entire structure.
</example>
</few_shot_examples>
<instructions>
Quote relevant details from the source material before generating prose.
Write the opening scene for the section on pricing psychology, drawing
on the Vulpine Creations experience. Direct, slightly self-deprecating.
400-500 words.
</instructions>Three techniques at work here. First, source material at the top, instructions at the bottom — with book-length context, this ordering produces up to thirty percent better output because the AI has the full picture before it starts generating. Second, three to five few-shot examples covering different section types (openings, analytical passages, reflective moments) so the AI has a concrete reference for whatever it needs to write. Third, the “quote before generating” instruction that anchors output in actual source material rather than letting the AI drift toward generic summaries.
I did not prompt “write chapter 7.” I worked section by section with specific direction for each part. The AI produced a first draft for each section. Some sections came out at eighty percent quality. Others needed heavy rewriting. On average, I estimate the first draft was about sixty-five percent usable as written, with the remaining thirty-five percent requiring significant human editing.
Stage 4: Editing and refinement. This was the most human-intensive stage and the most important.
AI’s role: I used a self-correction chain — three separate steps, each producing inspectable output. Step one: the draft from Stage 3. Step two: an editing agent reviewed the draft against structured evaluation criteria:
<evaluation_criteria>
<criterion name="consistency">Does this chapter use the same terminology
and frameworks as chapters 1 through N-1?</criterion>
<criterion name="factual_accuracy">Are all claims verifiable? Flag
statistics and named sources.</criterion>
<criterion name="voice_adherence">Does this match the voice examples,
or has it drifted toward generic prose?</criterion>
<criterion name="cross_reference">Do all references to other books
in the series point to correct content?</criterion>
</evaluation_criteria>Step three: a refinement pass applied the corrections. Each step produced visible output I could inspect. Why three steps instead of one? Because when something went wrong across six books running in parallel, I needed to see exactly where it went wrong. The review step flagged issues. The refinement step fixed them. I verified both.
My role: all substantive editing. Rewriting weak sections. Strengthening arguments. Adding personal stories that only I could tell. Removing anything that felt generic or obvious. This stage took longer than any other, which is exactly how it should be. The editing is where a book becomes a book instead of a long document.
Stage 5: Production. Formatting, layout, metadata, descriptions, and marketing materials.
AI’s role: generating formatted output, writing back-cover copy variations, creating chapter summaries, and producing promotional materials.
My role: final review and approval of everything. At this stage, the human time requirement drops significantly because the content decisions are already made.
The Daily Workflow
During the intensive production period, my typical day looked like this:
6:00-8:00 AM: Deep editing on whatever chapter was in Stage 4. This was my highest-energy work, so it got my freshest hours.
8:00-9:30 AM: Review AI-generated first drafts from the previous day. Mark sections that needed rewriting. Queue sections for the next round of AI drafting with specific feedback.
9:30-10:30 AM: Run AI workflows for Stage 1-2 work on chapters that were still being organized or outlined.
10:30 AM-12:00 PM: Client work and other business obligations.
1:00-3:00 PM: Write the sections that AI could not handle. Personal stories, nuanced arguments, sections requiring specialized expertise that the AI consistently got wrong. There were roughly two to three of these per chapter.
3:00-4:00 PM: Production tasks and cross-book consistency checks.
I did not work this schedule seven days a week. Five days was typical, with occasional Saturday morning sessions for editing backlogs. The timeline was intensive but not brutal, which mattered because quality degrades when the author is exhausted.
Tools and Costs
The tool stack was deliberately simple:
- Claude Pro subscription: EUR 20/month. My primary AI for drafting and editing.
- Claude API access: Roughly EUR 150/month for the automated workflows.
- n8n: EUR 20/month for workflow automation.
- Scrivener: One-time purchase. For manuscript organization.
- Google Docs: Free. For collaborative editing.
- Canva Pro: EUR 12/month. For cover design concepts (final covers done by a designer).
Total AI-related costs over the production period: approximately EUR 800. For six books. Compare that to hiring ghostwriters, editors, and research assistants, and the economics are striking.
But I want to be transparent: the real cost was my time. Over 500 hours of focused work. AI did not eliminate the work. It changed what the work was. Less time on research compilation and first-draft generation. More time on editing, refinement, and the creative decisions that make books worth reading.
What AI Handled Well and What It Did Not
AI excelled at:
- Organizing and clustering large volumes of source material
- Generating structured first drafts from detailed outlines and examples
- Maintaining consistent terminology and formatting across six books
- Identifying inconsistencies between chapters and between books
- Producing variations of marketing copy and descriptions
- Fact-checking claims and statistics referenced in the text
AI struggled with:
- Personal stories and anecdotes (it cannot know what I experienced)
- Nuanced arguments where the standard take is wrong (it tends toward consensus)
- Humor and voice (it gets close but not quite right without heavy example training)
- Knowing when to break rules for effect (it follows instructions too literally)
- Understanding what makes a specific book different from similar books
- Cutting content (it tends to add rather than subtract)
The struggled-with list is important. These are the areas where author investment is non-negotiable. A book written entirely by AI, even with excellent prompting, reads like a competent summary of existing ideas. What makes a book valuable is the author’s unique perspective, experience, and willingness to say things others do not. AI cannot provide that.
The Quality Question
I want to address this directly because it is the most common objection. “If AI helped write it, is it any good?”
Here is my answer: AI helped me produce drafts faster. Every sentence in the published books was reviewed, edited, or rewritten by me. The ideas are mine, drawn from my experience. The arguments are mine. The voice is mine, refined through hundreds of editing passes.
The analogy I use: a builder uses power tools instead of hand tools. Nobody questions whether the house is well-built because the builder used a nail gun instead of a hammer. The quality depends on the builder’s skill, judgment, and standards, not on whether the tools are manual or powered.
My standards for these books were the same as if I had written every word from scratch. The difference is that I could apply those standards across six books using AI-native methods instead of spending years on the same output. The AI raised my throughput ceiling without lowering my quality floor.
Reader feedback has validated this approach. Reviews consistently praise the practical, specific, no-fluff nature of the content, which is exactly what I optimized for during the editing stage.
What I Would Change
If I were starting the process today, three things would be different.
First, I would invest more time in Stage 2 (outlining) before starting Stage 3 (drafting). Some chapters went to draft too early and required structural rewrites that could have been avoided with a more thorough outline. The temptation to start producing pages is strong. Resist it.
Second, I would build the cross-book consistency checking system earlier. I built it in month three and had to retroactively apply it to the first two months of output. Starting with cross-book checks from day one would have saved at least twenty hours of rework.
Third, I would be more aggressive about the ship it ugly principle in the first drafts. I spent time polishing first drafts that were going to be rewritten anyway. Let the first draft be rough. Spend the polishing energy on the final version.
Can You Do This?
Probably, but with caveats.
If you have deep expertise in your subject, a clear structure in mind, and the discipline to edit rigorously, AI can help you produce a book (or multiple books) much faster than writing from scratch. The expertise and editing discipline are non-negotiable.
If you do not have deep expertise, AI will help you produce a book that sounds like every other book on the topic. It will be competent and forgettable. Do not do this. Write about what you actually know, and use AI to help you express it, not to generate knowledge you do not have.
If you want to try this process, start with one book, not six. Prove the workflow with a single project before scaling to multiple. And give yourself permission to take the time you need. Not everyone’s schedule or energy level supports an intensive pace.
Takeaways
-
AI is a production tool, not an authorship tool. It accelerates the mechanical parts of book creation while freeing you to focus on the intellectual and creative parts that make a book worth reading.
-
Invest heavily in outlining before drafting. A detailed outline with specific direction for each section produces dramatically better AI-generated first drafts.
-
Editing is where the book happens. Plan for the editing stage to take longer than any other. If AI-generated drafts need thirty-five percent rewriting, budget your time accordingly.
-
Maintain consistency systems for multi-book projects. Cross-book terminology checks, voice consistency reviews, and reference verification need to run continuously, not retroactively.
-
Your expertise is the non-negotiable input. AI can organize, draft, and format. It cannot provide the experience, opinions, and original thinking that make a book valuable. Those come from you.
