What it means

This report started with two questions and picked up a third along the way. How does Robert Greene write his books? What would it take to reproduce that process at scale with AI? And what could someone with bad intentions build with the same machine?

Fourteen chapters later, here are the answers, as short as they can honestly be.

How Greene writes#

Strip away the anecdotes and the method has six moves.

1. He reads a lot, and he reads slowly. Two to four hundred books per project, by his own counts, with a pen in hand. He prefers obscure and out-of-print books, follows footnotes from one book to the next, and drops weak books after a few pages (chapter 3).
2. He waits, then he writes cards. Weeks or months after a first reading, he goes back to his marks and copies what still matters onto handwritten cards, organised by theme and colour. A project ends with a box of two or three thousand of them (chapter 4).
3. The cards become chapters. He sorts the cards into themes, and the themes become chapters. For The Law of the Sublime, twenty-odd candidate chapters became twelve (chapter 5).
4. Every chapter opens with a story. Finding that story is, in his words, "the whole work." He wants high stakes, a turn and a surprise, then shrinks a long biography into a few pages and follows it with his interpretation (chapter 5).
5. He rewrites until it's alive. The first draft is "miserable", the fifth is readable, and around the twentieth it says what he wants. He cuts for boredom and refuses to preach (chapter 6).
6. He works alone, for years. Five or six years for The Laws of Human Nature, seven for the Sublime, the second one written by hand in more than a hundred notebooks after a stroke (chapter 7, chapter 8).

Underneath the six moves sit two things nobody can copy: the 18 years of odd jobs, failure and anger he says gave him something to write about (chapter 2), and the taste that picks one story out of a 600-page biography.

The numbers he gives move around from one interview to the next, and the report keeps them as ranges. The method doesn't move. He has described the same loop of read, wait, card, sort, story, rewrite for more than a decade, and he still describes it that way after the stroke.

What an AI version can take#

Line Greene's six moves up against what current models do well, and the split is clean.

The top half of that table is the job Greene gave his research assistants. Ryan Holiday's job included transcribing hundreds of hours of interview tapes and reading books Greene didn't want to read. Greene says most of the assistants after Holiday failed because they couldn't sit with boredom, had too much ego and had no sense of drama. A model has no trouble with the first two. The third is the open question of this report.

So the honest answer to "can AI reproduce Greene's process?" is that it can reproduce the research department he never managed to staff. It can't reproduce Greene.

What scales and what doesn't#

The cost chapter makes the point with arithmetic (chapter 10). Model costs barely register. Reading, carding, twenty drafts and every critic pass for a 500-page book come to about $700 with a sensible mix of models, or about $2,000 if everything runs on the most expensive one at list price. The books cost about $9,000, and the human time is worth more than both. You could run the pipeline a thousand times a year.

Two costs grow with every book, and neither is the model:

• The books themselves. Buying three to five hundred used and new books costs thousands of dollars per project. Scanning them adds time, even when OCR is cheap.
• Checking. A book of this kind carries thousands of factual sentences. Even when every sentence points to a card and every card points to a page, a person still has to read what the checkers flag and audit a sample of the rest. The verification chapter puts that at well over a hundred hours per book, and explains why the number doesn't shrink much with better models (chapter 13).

That's the core finding. The parts of the pipeline that scale freely are the ones that were never the bottleneck for a writer like Greene. The parts that don't scale are the ones that make a book trustworthy.

Three ways to build it#

There are three serious ways to put AI into a Greene-style book. They differ in where the human sits.

1. Full automation#

The model reads, cards, picks stories and writes, and a human skims the result. This is the cheapest version and the one the market already has too much of. Amazon limits self-publishers to three new titles a day because of it (chapter 14). The measured error rates in chapter 13 say what happens: some share of citations point nowhere, and some share of statements aren't supported by the sources they cite. Without gates, those errors ship.

The report doesn't recommend this version, and Greene would reject it outright.

2. Agents as researchers and critics, human as author#

This is the pipeline in chapter 11. Agents scout, mark, card, cluster and rank. The human signs off on the reading list and the outline, picks every story and decides the final text. Critics, including Vale, a fact checker and a bored-reader agent, review every draft.

Whether the model writes first drafts in this version is a dial, not a switch. At one end the drafter writes complete chapters and the human edits them. At the other end the human writes and the model only critiques. Greene's own line falls near the second end: he uses AI search for lookups and calls it "a tool like a pen," but says not to let it "compose paragraphs for you" (chapter 9).

3. Human drafts, model tags#

The third version keeps the writing entirely human and puts the model to work on provenance.

The author writes as usual, with cards made by hand or by the agents. When a draft is done, the model reads it sentence by sentence and does three things:

1. It proposes a card for every factual sentence. For each claim, it searches the card box and suggests the card that supports it, or marks the sentence as unsupported.
2. It checks each pairing. The quote on the card has to match the scanned page, and the sentence has to say what the quote says. This is the two-stage gate from chapter 13: a deterministic check first, then an entailment check.
3. It reports, it doesn't rewrite. The output is a list: sentences with a verified card, sentences with a weak match, sentences with no card at all. The author fixes them.

This version produces no AI prose at all, so it fits Greene's rule. It also attacks the failure that sank every incident in chapter 13: the quote nobody said, the source that doesn't exist. Every one of those was a provenance failure, not a style failure.

It's also the version with the clearest gap in existing software. The survey in chapter 12 found good tools for every neighbouring task, from OCR that keeps page structure to claim checkers and link checkers. Nobody yet builds cards from books you own, checks them against the scanned page, and enforces that every sentence of a manuscript resolves to one.

If I were building one piece of this report first, it would be that: a card verifier. Scan a book, read the printed page numbers, store each card with its exact quote and the words around it, and refuse any card whose quote isn't on its page. Then run a linter that fails the build when a factual sentence has no card. Everything else in the pipeline can come later.

What the dark machine teaches#

Chapter 14 imagined a fictional operator who runs the same pipeline on pirated libraries and throws away the gates. The chapter is a threat model, not a recipe, and it ends up making the same point as the cost chapter from the other side.

The operator saves on exactly the two costs that don't scale. No books to buy, because the corpus is pirated. No checking, because nobody reads the output carefully. What's left costs almost nothing, so the operator can produce in volume: knockoff books, fake biographies timed to real releases, summaries passed off as originals, science-flavoured nonfiction built from papers stripped of context.

The courts have already drawn the legal line. In Bartz v. Anthropic, Judge Alsup found that training on books that were bought and scanned was fair use, and that keeping a library of pirated copies was not. The settlement came to about $3,000 per book. For an operator who never settles anything, the bigger constraint is that a pirated corpus can't be cited. A card that points to a stolen scan isn't a provenance chain you can show anyone.

The defenses that work are mostly the ones this report recommends for honest builders anyway: provenance that readers can check, human sign-off that someone puts their name to, platform limits on volume, and certification schemes for human-written books. The cost of doing it properly is the protection.

What the report can't settle#

Some questions stayed open. Each is a real limit, not a formality.

• Can a model pick the right story? The story chapter shows how Greene chooses, and the pipeline chapter turns that into a rubric. Nobody has tested whether a model's top-ranked story matches what a writer like Greene would choose. That experiment is cheap to run and hasn't been done.
• Does the wait matter? Greene and Holiday both insist on a delay between reading and carding. The pipeline replaces it with a fresh-context second pass. Whether that captures what the delay does for a human reader is a guess.
• What happens to the writer? Greene's strongest argument against AI isn't about the book. It's about the person. He says the struggle with a hard paragraph built the "muscle" that lets him write at all. The report can measure costs and error rates. It can't measure that.
• The evidence about Greene is his own. Almost everything here about his method comes from interviews he gave, sometimes years after the fact, plus his assistants' accounts. The numbers move, and the report keeps them as ranges. A writer describing his own process isn't a controlled study.
• Some benchmarks are self-reported. Several tool comparisons in chapter 12 come from the projects themselves. The report flags them, but they haven't been reproduced.

Closing#

Ryan Holiday made the case against shortcuts more sharply than anyone in this report. Talking about the notecard system at a live event with Greene, he listed its flaws: that it is hard, that you can't scale it, that it takes forever, that your hand aches. Then he turned the list around. "That it's hard is the reason that it works," he said, and taking the difficulty away removes "all the value" (Greene and Holiday, live).

This report is a test of how far that holds. The answer it reaches is split. For the research, the difficulty can go: a model can read the books, copy the passages and check the citations, and the book gets better for it. For the writing, the difficulty is the point, and Greene's books are what that difficulty produced.

A pipeline built on that split would be faster than Greene and more careful than most of what's on the shelf. It still wouldn't write a Greene book. It might help someone write their own.

Sources#