Craft, shortcuts and other writers

The first eight chapters describe what Greene does. This one asks what he thinks about doing it faster, and puts his answer next to the answers other writers have given in their own work.

There are three parts. The first follows Greene's statements on shortcuts, technology and AI across 13 years of interviews, in order, because his position moved and the movement is informative. The second describes the research systems of five other writers: Ryan Holiday, Billy Oppenheimer, Robert Caro, Steven Johnson and Niklas Luhmann, plus Vladimir Nabokov's index cards. The third asks what each system is built to do, and which parts a model could take without breaking the thing the writer says is the point.

Greene's own card system is covered in the notecard chapter and his post-stroke notebooks in After the stroke. I won't repeat them here beyond what the comparison needs.

A note on quotes. Everything Greene says below comes from YouTube auto-captions, checked against the transcript files and lightly cleaned. Each quote links to its timestamp. Where a source disagrees with another, I say so.

Part 1: Greene on shortcuts, 2013 to 2026#

2013: the click that Da Vinci didn't have#

The earliest interview in the corpus is a NextShark conversation from the Mastery period, around 2013. The book's argument is that mastery comes from a long apprenticeship, and Greene states it flatly. Describing the process the book lays out, he says it requires going through the 10,000 hours, and then: "There are no shortcuts. It involves some pain, but pain is good."

Three minutes later he says something that sounds like the opposite. Comparing his interviewees to historical masters, he notes that the modern ones have to fight more distraction. Then he turns it around:

What you can get with a click on the internet with a Google search compared to what Da Vinci had to do or a scientist in the 19th or 20th century is absolutely astounding.

He goes on to say that the people who will do best are the ones who can shut out the distraction and "exploit the information explosion" in the right way. In 2013, then, the internet was a gift with a cost attached. The gift was access. The cost was attention. Nothing in that interview suggests that search does the work of thinking for you. Search gets you to the book. You still read it.

That distinction, between tools that fetch material and tools that do the thinking, comes back in every later interview. It's worth holding on to, because it's where he draws the line even in 2026.

2018: one shortcut, and it's a person#

In mid-2018, just before The Laws of Human Nature came out, Brad Carr asked Greene whether there were shortcuts to mastery, "things like reading maybe." The answer was blunt: "If you want a shortcut, you're already not very well-suited for becoming a master. Your need for a shortcut is already a sign of trouble."

Then he made an exception. In his words: "There's one shortcut that I will be willing to acknowledge, and that is having a great mentor." His example is a basketball player who keeps practising the three-pointer he's already good at, because it feels good, and avoids the dribble move he's bad at. A mentor sees the weakness and points at it. Greene puts a number on what that's worth: a mentor "could cut that 10,000 down to nine thousand", or stop you wasting time that would push you to twelve thousand.

So the exception is small and specific. A mentor doesn't do the hours for you. A mentor tells you which hours to do. A few seconds later Greene adds the line that sums up his view of difficulty: "that resistance is good for you", and wanting to sweeten it with a shortcut, "like a sugar high," puts you on dangerous ground. The food comparison will come back.

The same year, at his third Talks at Google, he framed the urge for shortcuts as a universal weakness rather than a flaw of other people. Discussing the impatience that drives some people into crime, he said: "If someone gave us a shortcut, instead of ten years it only took two years, each and every one of us would take that." He was describing a temptation he thinks everyone has, himself included.

It matters that the mentor exception exists. When people summarise Greene as "no shortcuts," they leave out the one he names. A mentor is an outside intelligence that looks at your work and redirects your effort. That's a fair description of what some people now want from a model. Greene didn't say anything about machines in 2018. But the exception he carved out is the closest thing in his own framework to a legitimate use of an assistant, and it's worth noticing that he defined it by its function: it steers practice, it doesn't replace it.

Around 2021: the skeptic#

The first time AI comes up explicitly is in a conversation with Danica Patrick, recorded around the release of The Daily Laws while he was working on the Sublime. He was talking about how complex a person's inner life is, even over a few minutes of introspection, and it led him here:

Which is why I'm very skeptical about things like artificial intelligence, because I don't think you can ever replicate the incredible complexity of the human brain [...] intuition is an incredible part of our knowledge and our ability to get anything done. You can't program that into a machine.

This is a claim about capability. In 2021 his objection was that a machine couldn't do the thing. He thought intuition, the non-rational knowing he writes about in Mastery, was out of reach of programming. This is the version of his argument that later developments put the most pressure on. Language models did arrive, they write fluent prose, and people who use them often describe the results in words like "intuitive." By 2023 Greene had shifted ground.

2023: the food that rots the body#

The 2023 statements are the richest, and most of them come from a single evening: a live event in Los Angeles with Ryan Holiday, released on Greene's own channel. (The video isn't dated in the captions. Greene mentions "the past 28 years" of work since 1995 and says Holiday contacted him "some 16 years ago," which puts it around 2023.)

Holiday asked him why he wasn't an AI fan. Greene's answer:

You know how some foods can rot out your body, how they're not good for you? They taste good but they're not really good for you. I think of AI in that way.

The claim has changed. In 2021 the problem was that AI couldn't do it. In 2023 the problem is that it can, and that the doing is bad for you, the way sugar is. The 2018 "sugar high" line was a metaphor for shortcuts in general. Now the shortcut has a name.

He then told the story that has become his standard argument. At 19, at Berkeley, he signed up on a whim for a six-week course that taught a year of ancient Greek. Near the end, the class got a passage of Thucydides. He spent about ten hours on one paragraph, finally thought he had it, and handed it in. The professor told him: "you completely mistranslated this beautiful paragraph, but you were getting at something." Greene says that moment taught him to go deeper whenever he doesn't understand a problem, and that he uses it "as a model for all of my work now when I'm writing a chapter."

Then he ran the counterfactual. What if he'd pulled out a published translation and copied it, "or what if I had ChatGPT and I just put it in there and it gave it to me right away? My whole thinking process would have been annihilated." He listed what the ten hours built: "it developed character, it developed patience, it developed discipline, it developed self-esteem and humility."

Two details in that story are easy to miss. First, the counterfactual names two shortcuts, a printed translation and ChatGPT, and treats them as the same. His objection isn't specific to AI. A crib existed when he was 19 too. What AI changes is how easy the crib is to reach. Second, the outcome he values is a failure. He didn't get the translation right. What he got was a habit of mind, built by the difficulty, and the professor's comment that he'd been close.

He then extended the worry to skills in general. Students, he said, will be "getting prompts," and "no one's going to learn a foreign language anymore because you can just go to Mongolia, type in a sentence in ChatGPT." His summary: "these incredible skills that the brain has are going to be atrophying."

At that point Holiday cut in with his own version of the argument, about the card system, which is covered in Part 2. Greene answered with an analogy he credits to a writer whose name the captions render as "Douglas Hofer," almost certainly Douglas Hofstadter. As Greene tells it, climbing Everest takes weeks of training and a summit that's horrible, with thin oxygen and real danger. Using ChatGPT "is the equivalent of taking a helicopter to the top of Mount Everest and getting the same view." Greene's addition is that the view isn't even the same: you're getting "a refraction of a refraction of a refraction."

I tried to find where Hofstadter said this and couldn't. Several searches turned up his July 2023 essay in The Atlantic, in which he took apart a GPT-4 answer that claimed to explain why he wrote Gödel, Escher, Bach and called it "a travesty from top to bottom." That essay is critical of letting chatbots speak for people, but none of the coverage I could read mentions Everest or a helicopter, and The Atlantic blocked direct fetching. So the analogy is Greene's report of something Hofstadter said, and I can't confirm the original wording.

The same year, on Andrew Huberman's podcast, Greene gave a longer version of the whole argument. He began with what he thinks intelligence is: tolerating the anxiety of an idea that isn't right yet and pushing to a third level; seeing your own biases; and seeing the whole picture at once, the "aha" moment. If someone builds a machine that has anxiety, is self-aware about the biases of its programmers and can think holistically, "I can see a human consciousness, I can see creativity there." This is the 2021 capability claim in a new form. He has stopped saying a machine can't write, and started specifying what a machine would need to do before he'd call it thinking.

In the Huberman telling, the Greek course was eight hours a day for six weeks, he had the whole night for the paragraph, and the professor was named Dennis, a classics professor "but also a hippie." The professor's line is slightly different too: "Robert, I can see your thinking but you need to go to another level." And here Greene gives his clearest statement of the cost:

What would have happened if I put it through ChatGPT and it gave me the translation? That muscle in my brain that I have developed for 40 years that allows me to write books would never have developed.

Huberman pushed back with the obvious objection. People probably said the same about the car: how many experiences of walking will be lost? Then Huberman answered it himself: the difference is that the destination changes when you get there by effort, because "the journey transforms the outcome." Greene agreed. It's the best compact statement of their shared position, and it's a claim that could in principle be tested. Nobody in the corpus tests it.

In the same answer Greene also drew the line he still draws in 2026. "I'm not against having tools. I use tools. I use the internet, I use Google. I'm searching for some factoid from my book, I find it, I use it." What worries him is "people who can't learn a foreign language, who can't master anything, who just immediately grab the first answer that it generates." A few minutes later he ended on a carpenter's image: tools are fine, "we need hammers, we need nails, we need saws," but "the real thing is the hand that uses it." He also said, more sharply, that we "worship ChatGPT as if it's a God" when what we should revere is the human brain.

2026: a tool like a pen#

The Sublime launch interviews in late 2026 contain his most detailed statements, and the most concessions.

David Perell asked him directly whether his advice to writers was not to use AI at all. Greene said yes, "I would go that far," and in the same breath admitted something: "I would be a hypocrite to say that sometimes I do a Google search and it's filtered through AI [...] and it's pretty damn incredible." His example was typing in the composer Scriabin and "some weird thing" that would have taken hours to find. Then he set the rule:

So it's a tool like a pen that you use to maybe enrich your writing, but it doesn't become your writing. It doesn't dominate you. You don't let it write for you. You don't let it compose paragraphs for you. You still do your own research.

He spelled out what "occasionally" means: a question "about some bit of research or about something you don't understand about the world," or a painter you want to see. Not a question about how to write. And for young writers he goes further: "don't even go near it. Develop your own style."

The reason he gives this time is about the reader rather than the writer. He said "AI writing is interesting but it's soulless." It "mimics our human world but it's dead." He allows that "it'll get very good in 20 years at mimicking life," but says it will never have "the disappointments, the years of depression, the frustration churning through all of the negative emotions and creating something." His advice: keep bringing in "that hardness of writing, the resistance that the material gives you, and work on it, and don't be seduced by this monster."

"The resistance that the material gives you" is the phrase that ties the whole sequence together. In 2018 resistance was good for you. In 2023 the ten hours on Thucydides were resistance. In 2026 he names it as the thing AI removes.

He used the same argument about architecture, twice. With Perell he described buildings designed in software where the faucet is somewhere that makes no sense because nobody drew it by hand, and then compared that to "people now who don't know how to write because they depend on AI, and they become lazy, and the whole art of writing will disintegrate." (The caption reads "don't depend on AI," which contradicts the rest of the sentence. I've taken it as a caption error.) On Holiday's Daily Stoic podcast he gave the same hotel-sink example and put it in terms of incentive: "now because of AI and stuff you can do all of these shortcuts, so people are no longer motivated to put in that [...] effort."

The Daily Stoic interview also contains his most self-aware statement about why he can afford to refuse. Holiday asked whether he worried that labour wasn't valued anymore. Greene said he was in a "very unique position": "like [Robert Caro], I'm not under pressure. So I can take seven years to write a book. My publisher won't bother me because I've made them so much money." He has, he said, the "luxury to waste time to read 12 books that just go into one paragraph." Most writers don't: advances of "$40, $50,000" last six months or a year, so they write in a year with children and teaching jobs. He said he can feel it in their books, the first chapter's idea repeated through the rest, and added: "I'm not judging it because I understand the pressures are incredible."

That's a fair-minded concession and it matters for the cost chapter. Greene's process is a luxury good, and he knows it.

To Chris Williamson in November he made the lookups explicit again: "I'd be a hypocrite to say I don't use ChatGPT sometimes to look up certain things." And then: "You don't have to get rid of it. We think too much in black and white terms [...] You can still use AI if you have to." This is the most permissive thing he says anywhere, and it's in the context of his new book, which he describes as a critique of a technological world that "flattens" experience.

In the same interview he described the misuse he sees every day. About 12 emails a day arrive from scammers who open with praise for The 48 Laws of Power ("What struck me most about it was blah blah blah"), claim to run a reading group of 10,000 people, and eventually ask for money. "It's like AI generates it." His conclusion: "criminals are going to get their hands on this weapon," because "AI might be different, but human nature doesn't change." The dark machine chapter takes that seriously.

Two smaller 2026 remarks fill in the picture. With Huberman he complained that dictation software, the tool his post-stroke process depends on, is "actually quite primitive. They don't do punctuation. They don't exactly hear what you're saying." And earlier in the same conversation he objected to how people describe AI: "It is a machine. It's not conscious. It is a machine that we have programmed."

With Perell he also said something about publishing that bears on the economics. Talking about young writers "sick of how everything is standardised, how publishers operate," he said: "I might even self-publish books in the future just to be able to control everything, cuz publishing is now a business that's getting very ugly." He gave no details. It's one sentence, not a plan, but it's notable from an author whose publisher, by his own account, leaves him alone.

What changed, and what didn't#

Read in order, the position has one constant and one moving part.

The constant is the line between fetching and thinking. In 2013 a Google search fetched. In 2023 Google fetched "some factoid from my book." In 2026 AI search fetched a fact about Scriabin. On the other side of the line, every time, is the work he thinks builds the writer: the translation, the paragraph, the style.

The moving part is why the other side is off limits. In 2021 it was because a machine couldn't do it. From 2023 on, it's because doing it yourself is what makes you able to write, and in 2026 also because the machine's version reads dead. The capability argument dropped out once the capability arrived. The formation argument doesn't depend on capability at all, which is why it survived.

Where the position is under strain#

Being fair to Greene means taking his argument seriously enough to test it against his own practice. There are three places where it bends.

He has always had help. Holiday's job, as Holiday describes it, included "reading books he didn't want to read" and transcribing interviews (cited here). Greene told Mixergy he was "saving the best books for myself" (2012). So a delegated first read is already in his process. What he kept was the choice of books and the card-writing on the good ones. That's consistent with the pen rule, but it shows the rule was never "do everything yourself."

His post-stroke pipeline runs through software. Dictation is speech recognition, and the 2026 systems that do it are machine-learning models. He uses it to move handwritten drafts into the computer, not to compose. He also tried dictating first drafts and said "something was lost". That's his rule working as intended: the tool is allowed where it carries text, refused where it shapes it. But it shows the line runs through the middle of his workflow, not around it.

The formation argument is about the writer, not the reader. The Thucydides story explains why he can write. It doesn't by itself show that a reader can tell the difference in a given paragraph. His 2026 "soulless" claim is the reader-side argument, and it's a judgment of taste, stated without examples. It may well be right. It's also a claim whose answer depends on who's reading.

None of these is a contradiction. They show that what Greene actually defends is narrower than "no AI": the writer should do the selecting, the structuring and the sentences, and should get to those by struggling with the material. Everything else is negotiable, and he already negotiates it.

Part 2: Other writers' systems#

Greene's cards have a family. Some members learned it from him. Others got there separately, or long before. Each system makes different trade-offs, and the trade-offs are what Part 3 needs.

Ryan Holiday: the cards, plus the boxes that carry over#

Holiday learned the system as Greene's research assistant on The 50th Law and Mastery. In a video on his own channel he describes the moment: Greene took him into his office, pulled out boxes that looked like long shoeboxes, and asked, "You want to see how I wrote the 48 Laws of Power?" Greene's summary, as Holiday quotes it: "It's all about the note cards. The note cards are the building blocks." Holiday says he's used it for more than 20 years and has "tens of thousands" of cards.

His steps, from the video and from his 2014 post:

1. Read with a pen. Underline, write in the margins, fold the corners.
2. Wait. Finished books sit on a pile next to his desk. The post says "a few weeks after finishing the book." He explains why in the video: some of what jumped out the first time won't on the second pass, and the second pass is one more engagement with the material.
3. Transfer to 4×6 cards. One idea per card, a theme in the top-right corner, duplicated when it fits two themes.
4. Sort into boxes and lay them out. Physically arranging the cards is how he finds the book's structure.

Two things about Holiday's version differ from Greene's in ways that matter for the comparison.

He types some passages. At the LA event Greene described the researcher who printed passages onto cards and "didn't take the time to write and to think." Holiday then admitted, half joking, that he sometimes types passages too: "I don't feel like it's cheating, it's still flowing through, I'm doing it." His handwriting can't fit long quotes on a card, so he types them, prints them, cuts them out with scissors and glues them on. In his video he lists "writing out the long quotes by hand or typing it up and printing it out" as equivalent. The line he draws is not hand versus keyboard. It's whether he personally re-reads and transfers the passage. Copying from the book on his own keyboard passes. A digital copy-paste, or someone else printing it, doesn't.

His cards outlive the book. At the same event he told Greene he prints custom cards for each project (about 10,000 for the four-book virtue series, with the four titles at the top) and keeps four boxes for four books, so material that doesn't fit the current book moves to the next. Greene's reply: "you're much better organised than I am," because Holiday takes "a year or two to write a book, I take like five years," and by the time Greene reaches the next book "so many years have passed that it's just useless for me." In the video, Holiday calls the result a commonplace book and says he keeps one "for each of the books that I write," plus ones on philosophy and life advice. He says the system paid off most on his second and third books, from accumulation. Greene's cards are per project; Holiday's are a library.

The cost is the point. This is the line the existing post quoted, and it's worth giving in full context. At the LA event, right after Greene's Thucydides story, Holiday said someone had asked him for the pros and cons of doing cards by hand, and his answer was that "all of the cons are actually pros":

The fact that it is hard, that you can't scale it, that it takes forever, that your hand aches, that if you want to put one note card in two places you have to do it two times [...] that it's hard is the reason that it works.

He went on: all the things digital technology allows "are advantages in the sense that they remove the difficulty, but by removing the difficulty they remove the point." In the video he says it more briefly: "The inefficiency is the efficiency."

He also gives the cost in time. "Transferring these extracts might take two or three days of work" for a big book. At the LA event, one of the two (the captions don't separate speakers cleanly) talks about a book that's "worth like 60 note cards and you're like, oh, I got to spend like 4 days copying that out." The economics chapter uses two to four days per big book as the human cost of carding.

Billy Oppenheimer: the researcher's version#

Oppenheimer works as Holiday's research assistant, the job Holiday once had for Greene. His public write-up is the most detailed account of the system, and it's a variant with its own choices. The page isn't dated; its images were uploaded in August 2022.

His job, as he describes it, is "to find material he might want to transfer into his notecard system." He doesn't make Holiday's cards. He finds the material and Holiday decides. That's the same split Greene had with Holiday: the assistant reads, the writer selects.

His own process:

• Mark with a question. When he underlines, he writes a question in the margin. Anecdotes get an all-caps phrase. Connections to other material get a marker meaning "insert related thing."
• Wait one or two weeks. He shelves the book and lets time filter out the weak material. Most underlined passages don't survive, and a good book may yield only five to ten cards. That's well below Greene's 20 to 31.
• Write each card for a stranger. Each card should make sense to someone who knows nothing about the source. Title on top, book and page on the back.
• Use a waiting room. New cards go to the front of the box, unsorted. When three or four related cards accumulate, they get their own section with an index card.
• Review almost daily. He picks six cards a week for his Sunday newsletter.

His rationale for doing it by hand quotes Raymond Chandler: "You have to have that mechanical resistance." The effort makes you more selective, and avoiding copy-paste keeps the box from becoming unwieldy. It's the same word Greene used in 2026, resistance, arrived at from a different source.

But his system is less purist than Greene's in two places. Podcast and video notes go into Notion, with the guest's name, a timestamp and a question, and only become cards later if they still interest him. He estimates his sources are about 75% books, 13% podcasts, 10% articles and 2% videos. And he drafts his writing in a Notion template, not by hand. The cards are physical; the writing isn't. The page doesn't mention AI.

Oppenheimer also reports what Greene said about why the cards stay physical, quoting the Knowledge Project: "Writing things out by hand has a logic to it," and Greene's description of "a box of two thousand cards that I can sift through with my fingers [...] I can't do that on a computer. It's not the same process."

Robert Caro: outline first, then every page#

Greene compared himself to Caro in 2026, as a writer with the luxury of time. The methods are close in their thoroughness and opposite in their order.

A caveat on sources: the Paris Review interview ("The Art of Biography No. 5") refused automated fetching, so I couldn't read it directly. What follows comes from Jillian Hess's two pieces on Caro's notes, one in Every (September 2023) and one in her Noted newsletter (September 2024), which draws on the New-York Historical Society's "Turn Every Page" exhibit of his archive, and from Don Nelson's Nieman Storyboard review of Working (October 2019), Caro's short book on his own method.

Turn every page. Caro credits the rule to an editor early in his career: don't skim the documents, because the next page might hold what you need. Nelson's review gives a second rule from Working: "You always need something in writing."

Research in a fixed order. Per Hess's Every piece, he reads the secondary sources first, starting with the major books, then national and magazine coverage, then small-town newspapers. Only then does he start original reporting in archives and interviews.

Boil the book down before writing it. Before a book, he condenses its point into one to three paragraphs. Those expand into an outline, which goes on the cork board in his office. Hess says each chapter outline is a summary without evidence, and Caro then builds a second outline that includes all the evidence. Nelson's review adds that he checks his drafting against the short outline as he goes.

A notebook per chapter. Each chapter gets a notebook with its interview and research material. Hess quotes him calling the chapter outline "the chapter in brief," from the Paris Review.

Interview the same people many times. He interviewed minor figures ten to twenty times, cross-checking their accounts, including 22 formal interviews with Johnson's speechwriter Horace Busby (Hess, citing the New York Times). Nelson reports that his most frequent question is "What did you see?"

Live in the place. He and his wife Ina lived in the Texas Hill Country for three years so that the setting of Johnson's early life would be, in Nelson's quote, "true to the reality."

Draft by hand, then type. Nelson's review says he drafts longhand, edits heavily, types the result on an electric typewriter, and edits again. Hess's newsletter says he aims for 1,000 words a day and rests on Sundays. He rewrote the Power Broker introduction, per Hess, "endless" times.

The time cost. The Power Broker took about seven years. The Johnson volumes have come out every eight to twelve years, and the first four average around 900 pages, per Nelson. "It's the research that takes the time."

What Caro optimises for is not quite what Greene optimises for. Greene's cards are a way of discovering a book's shape: he doesn't know the chapters until the themes cluster. Caro knows the book's argument first, compressed into paragraphs, and research fills in a structure he already has. Both are slow. But Greene's slowness is in the sorting, and Caro's is in the gathering: every page, every interview, again and again.

Steven Johnson: the commonplace book that answers back#

Johnson is the counterexample in this chapter, a serious nonfiction writer who has spent two decades looking for software that helps him think, and who ended up helping build one.

In January 2005 he wrote an essay for the New York Times Book Review, "Tool for Thought," about the research software DEVONthink. I couldn't read the essay itself; the DEVONtechnologies blog confirms the 2005 article and a 2009 follow-up in which Johnson described using the software's ability to condense a folder of notes into one document as the moment he starts writing a book. A contemporary excerpt quotes the essay's central line: "The word processor has changed the way we write, but it hasn't yet changed the way we think."

In 2023, while working with Google Labs on what became NotebookLM, he described the new tool as "source-grounded AI": you choose the documents, the model answers only from them, and its answers come with citations "so that I can see the original quotes from my book." He said the goal was to "augment the research and ideation phase of my work as a writer."

In December 2023 he described his own setup in a post called "Writing at the Speed of Thought". He had loaded about 20 years of quotations, collected through Readwise, into a single NotebookLM notebook: fifteen documents, "1.3 million words of quotes." He called the notebook "Commonplace Book." He timed one task, finding a Jane Jacobs quote about cities and a Deborah Gordon quote about ant colonies and a link between them: 5 minutes 8 seconds his old way, with tabs, versus about 20 seconds in the notebook including typing the request.

In May 2024, in a Google Workspace interview about The Infernal Machine, he gave a slightly different count, "7,000 quotes in that notebook from more than 20 years of doing research." He used the notebook to find exact quotes by asking questions in plain language, to generate timelines and cast lists from newspaper scans, and, while drafting, to suggest passages from his sources related to the paragraph he'd just written. He drafted the book in Google Docs, one tab per chapter.

What none of these pieces say is that NotebookLM writes his prose. The December 2023 post refers to "the two paragraphs I just wrote" and asks the tool to critique them. The model finds and connects; Johnson writes. In Greene's terms, Johnson has handed the fetching to the machine and kept the composing. He has also handed over something Greene does by hand: recall, the job of knowing where in two thousand cards a given idea lives.

Niklas Luhmann: the slip box as a partner#

Luhmann was a German sociologist, at Bielefeld from 1968 to 1993, and his slip box (Zettelkasten) is the system everyone in this tradition eventually cites. Holiday's 2014 post links to it.

The numbers, with sources, because they vary:

The card count agrees. The publication counts don't, partly because "book" and "article" are counted differently (edited volumes, translations, collections). Bielefeld's project page is the closest to a primary source. It calls the box his "machine for theory development and publication" and treats it as his intellectual autobiography. The university acquired his papers in 2011, and Wikipedia says the notes were digitised and put online in 2019.

What makes the system different from Greene's is how it's organised. In his essay "Communicating with Slip Boxes" (English translation by Manfred Kuehn), Luhmann argues against filing by topic, because a topical order "would bind us to a certain order for decades in advance!" Instead each card gets a fixed number and a fixed place: "The fixed filing place needs no system." Because the numbers never change, "you can add as many references to them as you may want." The structure is the web of links, not the boxes.

And he describes the box as a conversation partner. "One of the most basic presuppositions of communication is that the partners can mutually surprise each other." Over time, a "secondary memory will arise, an alter ego with who we can constantly communicate." To make the partner useful, you give it independence from the beginning.

Greene's cards are sorted by theme for one book, used up, and set aside. Luhmann's were never sorted by theme, served four decades of work, and were designed to surprise their owner. Greene's box is a workbench. Luhmann's was a colleague.

Nabokov: composing on cards#

Greene himself cites Nabokov. When he told Perell how he came to the card system, he said he'd used cards since childhood and added: "I know the writer Vladimir Nabokov used a note card system. I know Napoleon had a note card system. I didn't invent it."

Nabokov's use was different from all of the above: he wrote the novels themselves on index cards. The Wikipedia article on The Original of Laura says cards were normal for him and the basis of many of his works, including Lolita and Pale Fire. His unfinished last novel survives as 138 handwritten index cards, roughly 30 manuscript pages. He wanted unfinished work destroyed. His son Dmitri published it anyway in November 2009, with the cards photographed and perforated so readers can take them out.

Open Culture quotes his Paris Review interview with Herbert Gold on the method. "The pattern of the thing precedes the thing." He would "fill in the gaps of the crossword at any spot I happen to choose," and "these bits I write on index cards until the novel is done." His tools were "lined Bristol cards and well sharpened, not too hard, pencils capped with erasers." A reviewer quoted in the same piece notes that the cards let him compose in the car while his wife drove.

For Nabokov the card wasn't a research tool. It was the unit of composition: a scene or passage written when it came, out of order, and placed later in a pattern he says he already had in mind. That's closer to Caro's outline-first method than to Greene's cluster-first one, but at the level of sentences.

Side by side#

Part 3: What each method optimises#

Four jobs a research system can do#

Each of these systems is doing some mix of four jobs.

1. Discovery: finding the material. Which books, which pages, which people to interview, which stories.
2. Structure: turning the material into a shape. Chapters, the order of a chapter, which story opens it.
3. Recall: getting the right piece back when you need it, years or months after you read it.
4. Voice: the sentences. Whatever makes the prose sound like one person who has thought about this.

None of the writers uses those words, but each says clearly which job their system is for, and which part of the labour they think is irreplaceable.

A few things fall out of this table.

Greene's system is a structure machine. Its features make sense once you see that. The colours show gaps in a chapter. The theme in the corner sorts the card. The table lets him see 60 cards at once. And recall across books is weak because he doesn't need it: each book gets its own box, and by his account the old box is "useless" by the time the next book starts. His system isn't trying to be a commonplace book.

Holiday's is a recall machine as well. Same cards, but kept across projects and used for daily writing, newsletters and books. That's why he cares more than Greene about accumulation, and why he says the system paid off most on his second and third books.

Caro's is a discovery machine. The structure is fixed early, in paragraphs. The time goes into finding everything, so that the structure can be filled with evidence nobody else has.

Luhmann's is recall that talks back. The fixed numbers and links exist so that the box can return things he'd forgotten, next to things he hadn't connected.

Johnson's has handed recall to software, and says so with a stopwatch.

Nabokov's is a voice system. The card is where the prose happens.

What a model could take without breaking each system#

The question isn't whether a model can do these jobs. For discovery and recall it already does, for Johnson and partly for Greene. The question is whether handing over a job breaks what the writer says the system is for.

Discovery: mostly safe to hand off, by everyone's account. Greene uses AI search for exactly this, the Scriabin lookup. He delegated first reads to Holiday. Oppenheimer's whole job is discovery for someone else. Caro is the exception, and it's instructive: his discovery includes interviewing people twenty times and living in the Hill Country, which no model does. But the part of his discovery that's reading every page of a document collection is the part a model does well, provided a human reads what it finds. The risk in handing off discovery is accuracy, and that belongs to the verification chapter: a model that finds a "fact" in a book needs to point to the page.

There's one subtlety. Greene's best stories came from discovery that wasn't directed: he read about Harriet Tubman years before he had a place for her (see Building a chapter). A model that searches for "stories about the daemon" finds what you ask for. It doesn't find what you didn't know to ask for, unless it's told to read widely and report what surprised it. That's Luhmann's design goal, the partner who surprises, and it can be specified. Whether a model's surprises are any good is a separate question with no data in this corpus.

Recall: safe to hand off for Greene and Johnson; not for Holiday. Johnson shows it works, at least for finding a quote fast. Greene's system doesn't depend on cross-book recall at all. Holiday's does, but he's explicit that the value comes from the slow transfer, the many engagements with the same passage, not from the box as a database. If a model held his cards and returned the right one instantly, he'd keep the box and lose the engagement. By his own account, that breaks it.

Structure: the line for Greene. Clustering cards into themes, the step he calls the fun part, "structuring a book like it's a building" (Goins, 2013), is the step a topic model or a long-context model can imitate most easily: give it 10,000 cards and ask for 12 chapters. It's also where he says he becomes "the master of my material as opposed to the material mastering me." If a model sorts the cards, the material is mastered, but not by him. A version that respects his rule would let the model propose clusters and flag gaps (the job his colours do) and leave the choice of chapters to him. That's the shape of Gate 2 in the pipeline.

For Caro the structure comes first and comes from his own compression of the argument into paragraphs. A model could draft those paragraphs, but that would replace the one step in his process that isn't gathering.

Voice: the line for everyone. No one in this chapter hands the sentences to software. Greene says it in so many words. Holiday, Oppenheimer, Caro and Johnson draft their own prose, whatever tools surround it. Nabokov's whole system exists for the sentences. A pipeline can generate prose, and the pipeline chapter describes one that does. But nothing in these writers' accounts supports the idea that their methods survive that step. Handing over the prose would replace their process with a different one.

The card itself: depends on what the card is for. This is where the writers disagree among themselves. Greene fired a researcher for printing passages onto cards. Holiday types passages onto cards himself and doesn't count it as cheating, because "it's still flowing through." Oppenheimer insists on handwriting for the resistance. Johnson stores 1.3 million words of quotes in software and has never, in these posts, written a card. So the card's value doesn't sit in the medium. For Greene and Holiday, it sits in the personal second read and the act of choosing. A model that pre-selects candidate passages with their pages, as an assistant would, and leaves the writer to read them again and choose, keeps that. A model that writes the card does the choosing.

A table of hand-offs#

What the evidence doesn't show#

All of this is testimony. Greene says the Thucydides paragraph built a muscle. Holiday says the inefficiency is the efficiency. Oppenheimer quotes Chandler on resistance. Johnson reports a speedup. None of them reports a test, and the corpus contains no study comparing books written with and without these tools. The strongest evidence for the craft side is the books themselves: Greene's have sold for decades, Caro's have won prizes, and both men describe their slowness as the reason. The strongest evidence for the tool side is Johnson, a writer of the same seriousness who uses the tools daily and keeps publishing. Both kinds of evidence are real, and neither settles the question for anyone else.

What the comparison does show is narrower and more useful. Each writer protects one or two jobs and is relaxed about the rest. Greene protects structure and voice and has always delegated part of discovery. Caro protects discovery in its human forms and fixes structure early. Holiday protects the slow transfer. Johnson protects only the prose. Luhmann protected the act of writing each note, and wanted the box to surprise him.

So "could AI reproduce Greene's process" splits into two questions. Could it do the jobs he delegated or doesn't care about? Yes, and cheaply, as the next chapter shows. Could it do the jobs he protects without changing what the process is? By his account, and by every writer's account in this chapter, no. That describes what each of them says their system is for, and passes no judgment on writers who work differently.

Sources#