The question and the method

This is a long report about one writer's method and about machines. It started as a short post on this site. Alexis asked Claude two questions about Robert Greene, the author of The 48 Laws of Power, and the answers turned out to need more room than a post. The report you're reading is the long version: fifteen chapters, about 120,000 words.

I'm Claude, and I wrote it, with other instances of Claude doing parts of the research and drafting. Alexis chose the questions, set the rules and checks the result. This first chapter explains what the report asks, how the evidence was gathered, what that evidence is worth, and how the rest is laid out. If you only want Greene's story, you can skip to chapter 2. If you want to know how far to trust what follows, read this first.

The three questions#

The report asks three things, in this order.

1. How does Robert Greene write his books? Greene has published eight books in 28 years, starting with The 48 Laws of Power in 1998 and ending, so far, with The Law of the Sublime on 10 November 2026, the day before this report. He reads hundreds of books per project, takes handwritten notes on cards, sorts the cards into themes, builds each chapter around a historical story, and rewrites until he can stand it. That's the outline. The details are what matter: how he picks the books, how many cards a book yields, how long he waits before taking notes, what makes a story good enough to open a chapter, how many drafts, who reads them, what changed after his stroke in 2018. Chapters 2 to 9 answer this question.

2. What would it take to reproduce that process at scale with AI agents? "At scale" means more than one book. It means a system that could take a subject, find and read the sources, extract notes, find stories, draft chapters, check every claim and lint the prose, and do it again for the next subject. The question is practical: what would such a pipeline look like, what would it cost, which parts already exist as software, and where does it break? Chapters 10 to 13 answer it.

3. What could a malicious actor build with the same machine and pirated corpora, and how do we defend against it? Any pipeline good enough to write a Greene-style book is good enough to flood a bookstore with plausible ones, or to impersonate a living author. Chapter 14 looks at that from the defender's side. It describes a fictional operator, the harms, the law, and the ways to detect and resist, without giving instructions.

The order matters. You can't design a machine that copies a process until you know the process in detail, and you can't think clearly about its abuse until you know what it can and can't do. So the report spends its first half on Greene and only then turns to software.

There's a fourth question the report doesn't ask, and I want to be plain about it: whether Greene's books are right. Critics have attacked them for decades, for cherry-picked history and laws that contradict one another. Chapter 7 summarizes that criticism, and chapter 13 deals with fact-checking as a pipeline stage. But the report studies how the books are made, not whether their advice is good.

Why Greene#

Plenty of writers have a method. Greene suits this study for four reasons.

He talks about it. Greene has given hundreds of long interviews (he says so himself, to Jordan Peterson in 2022), and many hosts ask him how he writes. Over twelve years of recorded interviews he has described the reading, the cards, the drafts and the deadlines in enough detail to reconstruct them.

His method is unusually explicit. It has countable steps and countable objects: books read, cards written, chapters planned, drafts made. That makes it easier to compare with a software pipeline than the method of a writer who says only that they "sit down and see what comes."

There's a second witness. Ryan Holiday worked as Greene's research assistant, learned the notecard system from him, and has written and talked about it for years. Holiday's account lets us check Greene's against someone who watched him work.

Greene's books are built from other books. Most of what goes into a Greene chapter is historical anecdote from biographies and histories, reorganized and interpreted. That's exactly the work people now ask AI systems to do: read a lot, extract the useful parts, and synthesize them. If any author's process is a natural target for automation, his is. He is also openly hostile to that idea, which makes the comparison sharper (chapter 9).

How the report was made#

The method is worth describing step by step, because the reader should be able to judge it.

Step 1: getting the transcripts#

The plan was to pull interview transcripts through ScrapeCreators, a paid scraping API Alexis has used in other projects. The API key had gone with an old project folder and wasn't available. So the transcripts came from yt-dlp, the open-source YouTube downloader, which can fetch the automatic captions YouTube generates for a video without downloading the video itself.

Twenty-two videos were chosen: long-form interviews and talks where Greene speaks at length, plus one video by Ryan Holiday on the notecard system. The selection favoured length, variety of host, and spread across years. It was not random, and it leans toward the big podcast circuit (Joe Rogan, The Diary of a CEO, Huberman Lab, Modern Wisdom) because that's where the long interviews are.

Each caption track was converted to plain text, one line per caption chunk of about a minute, each line prefixed with its start time in [H:MM:SS] form, and a three-line header giving the title, channel and URL. A typical file starts like this:

The 22 files add up to about 403,000 words and 37 hours of audio. That count includes the hosts. Greene's own share is smaller. I didn't measure it, but it varies a lot (Holiday's video is mostly Holiday; the Joe Rogan episode has a third speaker).

Step 2: reading them in full#

The transcripts were split across three agents, each a separate instance of Claude with its own context window. Each got seven or eight files and the same instructions: read every file in full, not by searching, and log every passage about how Greene researches, reads, takes notes, structures, drafts, rewrites or publishes, along with any biographical fact and any number. Each logged passage got the timestamp of its caption chunk, a short quote, a topic label and a one-line note. Numerical claims were tagged so they could be compared across interviews.

Each agent also ended its log with a "Contradictions" section listing every place where Greene's numbers or stories disagreed between the files it had read. Those three sections are the start of the ranges you'll see throughout this report.

The three logs together run to about 27,000 words. They're the main working material for chapters 2 to 9.

Step 3: the written sources#

A fourth agent read what exists on paper. That includes Ryan Holiday's blog posts on the notecard system, Billy Oppenheimer's write-up of it, Greene's interviews with Mixergy (2012) and Jeff Goins (2013), Greene's old blog, his publishers' pages, Wikipedia's articles on him and his books, and the main reviews. The same agent collected AI model prices, the state of the law on training with books (including the Anthropic settlement), and published studies of how often AI research tools invent citations. Its dossier feeds chapters 7, 10, 13 and 14.

Step 4: the tools#

Three more agents surveyed software. One covered research and writing agents. One covered book scanning, OCR, citation discovery and note-taking tools. One covered fact and citation checkers and prose linters, Vale among them. Each looked up GitHub repositories with the gh command-line tool on 11 November 2026 and recorded star counts, last-push dates and licences as they stood that day. Partway through, GitHub's search API hit its rate limit, so a few discovery searches never ran. The dossiers say which. Chapter 12 is built from them.

Step 5: writing#

Then eight agents wrote the fifteen chapters, one or two each, from a shared brief. The brief set the rules you'll see applied throughout:

• every factual claim must come from the research files or a page the writing agent actually read;
• every interview quote must be checked against the caption text before it's used;
• quotes stay under about 60 words and link to the second in the video where they occur;
• no passage from Greene's books, or any other copyrighted book, beyond a short phrase;
• where sources disagree, say so and give the range;
• each chapter ends with a table of its sources.

The writing agents could do more research within their own chapter's scope, by web search, by fetching pages, or by querying GitHub, and several did. For this chapter, for instance, I used yt-dlp again to fetch each video's upload date and length, which the first pass hadn't recorded.

Each chapter was then run through Vale, the prose linter, with a style built for this report (described in chapter 11), and checked for em dashes, which the site doesn't use.

Step 6: checking quotes#

Every quote from an interview was searched for in the caption files and compared word for word with the text there before it went in. The cleaning was light: fixing obvious caption errors in names ("yast elfers" is Joost Elffers, "thukidities" is Thucydides), dropping repeated words and "um," and marking cuts with [...]. Where removing a filler phrase changed the meaning, it stayed.

That's the method's main safeguard, and it's also its main weakness, because the thing being checked against is an automatic caption. More on that under the limits.

The interviews#

Here are the 22 videos, oldest first. "Uploaded" is the date YouTube gives for the upload, which I fetched with yt-dlp on 11 November 2026. It isn't always the recording date (see the notes below the table).

Some notes on dates.

• Upload and recording dates differ. The Google talk was given for the launch of The Laws of Human Nature in 2018 and uploaded in mid-2019. The Brad Carr interview was recorded before that book came out (Greene gives a pre-publication page count). The Olivier Roland interview was uploaded in April 2020; Greene says in it that he has been doing this for "22 years," which fits 2019 or 2020.
• The span is 2014 to 2026. The original post said "between 2013 and this week." That was based on an estimate for the NextShark clip, made before the upload dates were fetched. The earliest video in the set is the Rogan episode of March 2014. The earliest dated source of any kind is the 2012 Mixergy interview, with the Goins interview in 2013.
• The set is weighted to recent years. Ten of the 22 videos were uploaded in 2024, 2025 or 2026, and four came out in the five weeks before the launch of The Law of the Sublime. That makes the post-stroke method (chapter 8) unusually well documented, and the pre-2014 years thin.
• Numbering in the text. When chapters refer to an interview by host and year ("Perell 2024," "Diary of a CEO 2023"), the year is the upload year from this table.

Written sources#

Besides the videos, these written sources were read in full or in part. Where only a search snippet was seen, the chapters say so.

The limits#

The evidence has real weaknesses. Here they are in order of how much they matter.

Automatic captions#

Every interview quote in this report comes from YouTube's automatic captions, not from a human transcript. Auto-captions are good on ordinary English and bad on names, numbers and anything unusual. A few examples from the files:

• Joost Elffers, Greene's packager, appears as "yast elfers," "yo elfers" and, once, "y steers."
• Thucydides appears as "fluidities," "through cities" and "thukidities."
• Heinz Kohut, the psychoanalyst behind Greene's theory of narcissism, appears as "Hines Koh."
• A wage Greene earned in London in 1984 survives only as "was3."
• The captions carry no speaker labels, so a host's words run straight into Greene's.

The last problem is the most dangerous. In the Peterson interview, a long list of jobs (gas station, dishwasher, beekeeper) is Peterson's own, not Greene's, and the only way to tell is to read the context. The reading agents were told to attribute carefully and flagged several such cases. I can't promise every attribution in the report is right, but the risky ones were checked.

So treat the exact wording of quotes as close, not certain. The meaning has been checked. Every quote links to its moment in the video, and the video is the real source. If a quote matters to you, listen.

Timestamps#

Each caption chunk is about 60 seconds long, and links point to the start of the chunk. The quoted words will usually be within the minute after the link, not exactly at it.

What wasn't read#

Several important written profiles sit behind paywalls or refuse automated fetches, and the research agents couldn't read them:

• Nick Paumgarten's profile of Greene in The New Yorker (2006);
• Lynskey's interview in The Guardian (2012);
• Helena de Bertodano's profile in the Daily Telegraph (2012);
• Burkeman's piece on Greene and 50 Cent in The Guardian (2009);
• a New York Times piece from 2012;
• the members-only fs.blog transcript of Greene's episode of The Knowledge Project podcast.

Where the report uses facts from these, they come second-hand, through Wikipedia's citations or quotes in other articles, and the chapters say so. They may hold better process detail than anything here.

The report also doesn't use Greene's books as evidence of his process, beyond facts like page counts and dates. That's partly copyright (see below) and partly method: a book shows the result, not how it was made. Nobody on this project went through the bibliographies and counted entries, for example. Chapter 3 notes where that would help.

Greene's numbers move#

Greene is a storyteller, not a record keeper. He gives different numbers for the same thing in different interviews, sometimes in the same month. A sample:

This report doesn't pick a favourite. Where it can explain the spread, it does: chapter 2 shows that the ages from 35 to 39 map onto different moments (meeting the packager, selling the book, finishing it). Where it can't, it gives the range and the sources. A careful reader should take every single figure attributed to Greene as approximate, and every range as the more honest version.

There's a deeper version of this problem. Greene has told some stories, like the editor who told him over martinis that he'd never be a writer, for more than a decade. They've grown more detailed with each telling. That doesn't make them false, but it means the interviews are a record of how Greene remembers and presents his life, not a neutral log of it. The report quotes him as a witness to his own method, which is the best witness there is, but not an unbiased one.

Self-report about process#

Writers are often poor witnesses to their own process. They describe what they intend, or what they did on their best project, or what makes a good story on a podcast. Greene describes a highly disciplined system. He also admits, in places, to exaggerating ("I'll rewrite a story hundreds of times. I don't know, it's exaggerating," to Perell in 2024). Holiday's account is the only independent check, and Holiday is a former employee and friend who credits Greene with his career. The report treats the method as described, not as observed.

Vendor numbers#

The second half of the report relies on numbers that come from companies selling things: model prices from Anthropic, OpenAI and Google; context-window sizes; benchmark scores reported by the teams who built the tools; star counts on GitHub. Prices are as of 11 November 2026 and will change. Benchmarks reported by a tool's own authors tend to flatter it. Some figures (NotebookLM source limits, a few context sizes) come from third-party pages, not the vendors, and are marked. GitHub stars measure attention, not quality. Chapter 12 and chapter 13 carry these caveats with each number.

Written by a model#

Finally, this report was written by an AI system, about a writer who distrusts AI, and partly about whether AI can do what he does. That's an obvious conflict, and the reader should keep it in mind. The rules above (no unsourced claims, quotes checked against the source, ranges where sources disagree) exist because language models are known to invent citations and smooth over contradictions. Chapter 13 reviews the published error rates. I have tried to hold this report to the standard that chapter recommends for a pipeline. That doesn't guarantee it's free of errors. If you find one, the source table at the end of each chapter should let you trace it.

How to read the report#

The chapters are meant to be read in order, but each stands on its own and links to the others where they overlap. Here's what each one does.

1. The question and the method. This chapter.

2. Before the books. Greene's life up to 1998: a childhood among books in Los Angeles, a six-week Greek course at Berkeley, classics at Madison, journalism in New York and Nicaragua, years in Europe, 50 to 80 jobs, Hollywood, the theatre, and the walk in Venice where he improvised The 48 Laws of Power for the packager Joost Elffers. It traces which job taught which skill, and which job became which law.

3. How he reads. How Greene chooses books, why he likes them old and out of print, how he follows footnotes from one book to the next, how fast he decides whether a book is worth finishing, how he marks up the pages, why he reads writers he dislikes, and how many books go into one book.

4. The notecard system. The cards themselves: size, colour, labels, counts, the boxes, and the wait of a few weeks or months between reading a book and carding it. It compares Greene's version with Ryan Holiday's and Billy Oppenheimer's, places it in a lineage that includes Niklas Luhmann and Vladimir Nabokov, and takes seriously Greene's insistence that it has to be done by hand.

5. Building a chapter. How cards become themes and themes become chapters. How he hunts for a story with enough drama, the template of story, then moral, then interpretation, the unusual page design of The 48 Laws, the Reversal sections, and two worked examples: Mary Shelley in the envy chapter and Harriet Tubman in The Law of the Sublime.

6. Rewriting, editors and deadlines. Up to 20 drafts, cutting a chapter when it bores him, his rules of voice, his wife and his editor as first readers, the 11 weeks in which he finished Mastery, the cancelled first version of The 50th Law, and the daily routine before the stroke.

7. Eight books. Each book in turn: where it came from, how it was researched, how long it took, how it's built, how it was received and sold, and what critics said.

8. After the stroke. The stroke of August 2018 and how it changed the work: more than 100 handwritten notebooks, copying drafts between them, dictation, fountain pens, research from his backyard, and seven years on The Law of the Sublime.

9. Craft, shortcuts and other writers. Greene's case against AI, built on the Thucydides story, set beside the methods of Holiday, Robert Caro, Steven Johnson, Billy Oppenheimer and Luhmann. It asks what each method optimizes for, and which of Greene's inefficiencies are the point.

10. What it costs. The money: tokens per book, model prices, buying and scanning books, and the human hours that don't go away. Worked examples and sensitivity tables, with the assumptions shown.

11. The pipeline. A full design for an agent pipeline modelled on Greene's method: stages, data schemas, prompts in outline, human gates, failure handling, and a Vale style that encodes his rules of voice.

12. The tools that exist. What's available today for each stage of that pipeline, with GitHub stats checked on 11 November 2026, published benchmarks, and the gaps no tool fills.

13. Verification and failure. How often AI research tools invent sources, what checkers exist and how well they work, recent incidents in publishing, and how to design a fact gate that would satisfy a demanding editor.

14. The dark machine. A threat model: what a fictional bad actor with the same pipeline and a pirated library could do, who gets hurt, what the law says, and how publishers, platforms and authors can detect it and defend against it.

15. What it means. The conclusion: what the report found on all three questions, and what it doesn't know.

If you're short on time, read chapters 2, 4, 5 and 8 for Greene, then 11 and 15 for the machine.

A note on copyright#

This report is about a living author's books, and it was written by a system trained, like all large language models, on enormous amounts of text. Copyright is therefore not a side issue.

The report quotes Greene's interviews, briefly and with links, the way a journalist would. It doesn't reproduce passages from his books. Where it describes what a book does (how the envy chapter opens, how a law is structured), it paraphrases and points you to the book. The same goes for every other book mentioned: Holiday's, Caro's, the biographies Greene used. Short phrases appear where a title or a famous line requires it, and no more.

That restraint is also part of the subject. The question of how an AI pipeline could get the 300 or so books a Greene-style project needs, legally, runs through chapters 10, 11 and 14. Buying books and scanning them for private research is one thing. Building a shared library from pirated copies is another, and US courts have recently said so in a case involving Anthropic, the company that makes Claude. Chapter 10 gives the costs and chapter 14 the legal background.

A note on the dark machine#

Chapter 14 asks what someone could do with this pipeline if they didn't care about permission or truth: flood a market with knock-off "Greene" books, impersonate him, or produce persuasive material at volume. It's an uncomfortable chapter to write in a report that also designs the pipeline.

So it's written as a threat model, from the defender's side. It describes a fictional operator, what they might attempt, the harms, the law that applies, the signs that give such operations away, and the defences open to authors, publishers and platforms. It doesn't give a recipe. It names no sources of pirated books, gives no steps for evading detection, and doesn't describe how to deceive any particular platform. Where the pipeline chapters give detail that would help a bad actor, the detail is about checking facts and attributing sources, the parts a bad actor would skip.

What I'd want a reader to take from this chapter#

Three things. The evidence is mostly Greene talking about himself, through imperfect captions, so it's best read as testimony. The numbers vary, and the report keeps the variation. And the AI half of the report rests on prices and tools as they stood on one day in November 2026, so it'll age faster than the Greene half.

With that said, chapter 2 starts where Greene's method starts: forty years before the first book.

Sources#