Most people assume the order is: first you think, then you write it down. The thought is finished somewhere inside, and writing is the transcript. Anyone who has ever sat in front of an empty page knows this is wrong, and yet the belief survives — mostly because it lets us blame the page instead of the thought.

Heinrich von Kleist described the real order in 1805, in an essay whose title is already the argument: Über die allmähliche Verfertigung der Gedanken beim Reden — on the gradual completion of thoughts while speaking. His advice was to start talking to someone about a problem you have not solved, and to trust that the solution will assemble itself somewhere in the middle of the sentence. Not because your listener helps. Because the sentence forces you to commit.

There is a proof of this hiding inside the most famous sentence in philosophy. Descartes never wrote «I doubt, therefore I think, therefore I am.» He wrote «I think, therefore I am», and left the doubting in the pages before it, where the reader had to go and find it. The three‑part version was written by Antoine Léonard Thomas in 1765, in a prize essay praising a man who had been dead for 115 years. Today it is quoted as Descartes almost everywhere, including by people who have read him. The thought was finished outside his head, by a stranger, in a text he never saw. It took 128 years, but the loop closed.

Descartes had the doubt, he just kept it in the paragraph before. When he asks in the second Meditation what a thinking thing actually is, his answer opens with a thing that doubts, and only then understands, affirms, denies. The doubting came first for him too. What he did not have was a reader’s sentence for it, and a reader wrote one.

Writing does the same thing, only slower and with a record. Media theory has a blunt word for this: tools are not containers for thought, they are part of it. Andy Clark and David Chalmers made the philosophical case in 1998 — the notebook in your pocket is not a description of your memory, it is your memory, just stored outside the skull. Niklas Luhmann said something similar about his famous slip box, and he meant it literally: he described it as a conversation partner, not as an archive. Sönke Ahrens turned that practice into a manual, and the manual’s real subject is not filing. It is the loop.

The loop is the whole method

Write it down. Read it again a week later, when you are no longer the person who wrote it. Notice that it does not say what you meant. Rewrite it. Notice that the rewrite contradicts a note from three months ago. Follow that contradiction, because it is the most interesting thing you own.

None of that happens in your head. It happens because a thought was made into an object you could come back to and disagree with. And it only works if you allow the first version to be bad.

The first draft of anything is shit.

That line is attributed to Hemingway, and I use it constantly — so it seemed fair to check it. The source is Arnold Samuelson’s memoir With Hemingway, published in 1984 about a conversation in 1934. Fifty years between the sentence and its publication is a long time for a memory to stay exact. A letter Hemingway wrote to Fitzgerald in 1934 makes the attribution plausible — «I write one page of masterpiece to ninety one pages of shit» — but plausible is not verified. I am keeping the quote and the caveat, which is, I think, exactly the kind of small revision the method is about.

Where AI fits, and where it doesn’t

A prompt is a draft. That is the useful way to see it, and it is why prompting rewards the same discipline: you do not know what you want until you have written a version that is wrong.

The risk is not that a model writes badly. It is that it writes adequately, immediately, and removes the friction that made the loop work. If the first draft is generated, there is nothing of yours to argue with — you skip straight to editing something average that you never disagreed with in the first place. The productive move is the opposite: write your own bad version first, then hand it over to be attacked. Ask for the counter-argument, not the polish. Use the model as a mirror, not as a ghostwriter.

This is the same idea as everywhere else on this site. Raw force throws nobody; the energy needs direction. A model that finishes your sentences is force without direction, and it feels like progress right up to the moment someone asks you what you actually meant.

The part I am genuinely worried about

A generation is now growing up that may never write by hand at any length. There is research here, and it is worth being careful with it. Van der Meer and Van der Weel measured 36 students between 18 and 29 with a 256-channel EEG and found widespread theta and alpha connectivity while handwriting that simply was not there while typing — patterns associated with memory formation. The authors are honest about the limit: they did not show that this produces better learning outcomes. So the honest version of the argument is not «science says pen beats keyboard».

The honest version is simpler and harder to dismiss. You cannot write by hand as fast as you can type, and you cannot type as fast as a model can generate. Each step removes a constraint, and every one of those constraints was doing work. Slowness forces compression, compression forces choices, and choices are the thing we later call thinking. Take the slowness away and you keep the output while quietly losing the process that used to justify it.

I have no interest in nostalgia about paper. I write most things in a text editor and I use models daily. But I still carry a dotted notebook, and it is not decoration — it is the only place where I am slow enough to notice that I was wrong. The grid of dots behind this page comes from that notebook. It seemed like the right thing to build the whole site on.

Sources: Kleist, Über die allmähliche Verfertigung der Gedanken beim Reden (1805) · Descartes, Discours de la méthode (1637) and Meditationes (1641); the three‑part version is Antoine Léonard Thomas, Éloge de René Descartes (1765) · Clark & Chalmers, The Extended Mind (1998) · Ahrens, How to Take Smart Notes (2017) · Samuelson, With Hemingway (1984), via Quote Investigator · Van der Meer & Van der Weel, Frontiers in Psychology (2024).