// writing

2026-06-06 // 12 min // ai, cognition, writing

a tool for thinking

I had a problem last week with too many moving parts to keep in my head at once. So I put all of it on the page, worked it for an hour, got somewhere, and stopped. The next morning I opened the file and I was back inside the problem in a minute, not from memory, from the page.

The page did two things for me there. It held more than I could keep in my head, and it kept the state of my thinking so I could walk away and come back to it. That’s writing doing the thing it’s genuinely good for, which isn’t being the thinking. It’s being a tool I think with.

I said in the brain drives the tool that the judgment stays yours. The thinking is that judgment. This is about one of the tools you do it with, and what goes missing when you let an LLM do it for you.

writing is a tool for thinking

Let me get the claim right, because it’s easy to overstate and the strong version is wrong. Writing isn’t the thinking. It’s a tool you think with. Two plain tells. You think hard all the time without writing a word, in the shower, mid-argument, on a walk with your hands in your pockets. And writing does plenty of things that have nothing to do with thinking, like telling someone what you already worked out.

So the frame I mean is the modest one out of embodied and extended cognition: the mind leans on the body and the world to do its work, and writing is one of the things it leans on. Not the grand version where the page becomes a literal piece of your mind. The tool reading is the one that survives contact, and it’s the one I’m making.

You can see the gap most clearly when a thought shows up whole before you’ve written anything down. It was already there, finished, waiting to be transcribed.¹ If writing were the thinking, there would be nothing sitting there to transcribe.

What the tool does, mechanically, is make the thinking cheaper to run. Kirsh and Maglio watched expert Tetris players spin the falling pieces on the screen instead of rotating them in their heads, because doing it out in the world is quicker and surer. They called it epistemic action: a physical move whose job is to cut what you have to hold, the steps you have to run, the mistakes you’d make running them.² Writing is that move, aimed at thought.

The tool does two jobs, really. It holds what won’t fit in your head, and it keeps what you’d otherwise have to carry. Take them one at a time.

more than you can hold

Start with the holding. Working memory is small, famously so. You can keep a handful of things active at once, and a hard problem has many more than a handful of parts, each one bearing on the others. Try to keep all of it live and your effort goes into not dropping pieces instead of into thinking about them.

So you put them down. On the page they hold still. You can look hard at one part while the rest wait in plain sight, then the next, instead of juggling the whole set and losing half on every pass. You end up working on far more than your head could hold, because most of it is sitting on the page and only a little is in mind at any moment.

The name for this is cognitive offloading, using something outside the head to cut what the head has to do, and Clark and Chalmers made it concrete with Otto, who keeps in a notebook what the rest of us keep in memory and works the problem there on the page.³ The notebook carries the load. Otto does the thinking against it.

It’s also how I get further than I could in my head alone. A thought too big to hold together all at once, I can build in pieces on the page, one part at a time, until the whole thing stands up and I can finally see it.

You know the feeling. A refactor with eight constraints that each move the others. A decision with a dozen considerations that won’t sit still. Six premises you can’t keep straight long enough to see whether they actually connect. None of it fits in your head. On the page it fits, and once it fits you can work it.

The buffer holds the problem. The thinking still happens in you, just with more room than your head alone would give it.

somewhere to leave it

Now the part I lean on hardest. The page doesn’t only hold the parts of a problem. It holds the state of the thinking: the paths I tried, the ones that died, what’s still open, what I’d already decided and why. The whole position, not just the pieces.

That’s what lets me stop. I can put a hard problem down, sleep on it, and come back the next day by reading myself back into it off the page, instead of rebuilding it from memory, which I can’t do well anyway. What I come back to is the position I left, not a lossy reconstruction of it.

There’s a real result under this. Sparrow and her colleagues found that when people expect to have something stored where they can get back to it, they stop remembering the thing and start remembering where to find it. The store becomes what psychologists call transactive memory, where the information lives “outside ourselves.”⁴ That’s the trade I’m making on purpose. I don’t carry the state of a hard problem in my head. I carry a pointer to the page that holds it.

This is what makes writing a tool for thinking across time, not only inside one sitting. Working memory wipes the moment you look away. The page doesn’t. So the thinking can run for days, an entry at a time, because the state survives the gaps between.

The decision I leave open for a week, adding a line whenever something occurs to me, is built on a week of thinking by the time I make the call. Held in my head, it would have been built on whatever I could still remember on the last afternoon.

the work you didn’t do

Now put the LLM in. You hand it the problem, or a rough start on it, and back comes the finished, arranged result. It ran the working-through. Every move across the buffer that would have been yours happened somewhere you can’t see, and what reaches you is the clean, finished result with none of the working that produced it.

For anything you hadn’t already worked out, you’re holding the result without the understanding, because you never loaded the problem into the buffer and worked it. The working-through your own writing would have carried, the LLM carried instead.

There’s a second loss, and it’s particular to the saved state. What the LLM hands back is an artifact, a clean finished thing. It is not a state of your own thinking you can climb back into. No half-built moves, no dead ends, no trace of how it got there, because none of that happened in you. So there’s nothing of yours to pick back up tomorrow. The save file belongs to someone else.

Let me keep this honest with a bound. If you already worked the problem out, in your head, on a walk, last week, then having the LLM write it up costs you nothing real. There was no thinking left to do, and you’re using it to type. The loss is the problem you hadn’t worked through yet, where moving through it on the page was how you would have.

And none of this needs the page to be literally part of your mind. Adams and Aizawa are right that coupling a tool to a job doesn’t make the tool a piece of the worker.⁵ That’s the point, not a problem. The page is a tool, not a piece of you, so when the LLM runs the working-through, your mind stays exactly where it was. It never ran.

the artifact isn’t the understanding

Here’s where the evidence cuts against me at first, so let me take it head on. Hand the writing to an LLM and the writing gets better. Noy and Zhang ran the study: people doing real writing tasks, half of them with the LLM, and quality went up by about 0.4 of a standard deviation while time went down.⁶ On the page, measured, the work improves.

Concede that fully, then look at what the measure missed. It graded the page. It didn’t grade the person. And the tell is sitting in the same study: 68% of them submitted the LLM’s output unedited.⁶ Whatever the quality score says, most of them did none of the working-through.

The cleanest proof that the page and the understanding come apart is old and simple. The generation effect: give people material to read, or have them produce the very same material themselves, and the ones who produced it remember and understand it better. Identical words. The only difference is who did the work of making them.⁷ So the gain was never in the words. It was in the producing.

It happens with single words, too. An LLM will hand you the exact specialist term, and you’ll set it in a sentence and sound like you know the field. Feynman’s father taught him a bird’s name in five languages and then told him he now knew nothing at all about the bird.⁸ The term shows up finished, the understanding that earns it skipped. The same loss, one word wide.

what’s left

So, back to the start. the brain drives the tool said the judgment is the part you can’t hand off. The thinking is that judgment, and writing is one of the tools you do it with. Let the LLM do the thinking and the judgment has nothing left to form on.

It’s the same loss I wrote about in the reps you stop taking, shrunk to fit one problem in one sitting instead of a whole career.

Use the LLM for the cold case: writing up what you already hold, the breadth, the parts you’ve worked a hundred times. When you haven’t worked the problem through yet, do that part on the page yourself. That’s not where the thinking happens to sit. That’s the thinking.

It’s why these posts exist, too. They’re tools for my own thinking, the end product of me chasing something until I understand it. I’m not building an audience or a funnel with them, and I’m not trying to move what you think about anvilwright. If one of them is useful to you, good. That was never the job.