This piece follows directly on from my article "Flash Card". If you haven't read it yet, I'd start there. It explains what a pillar card is, what sets it apart from an atomic note, and why it has to stay rare. Here, I want to show how I keep those cards alive.

I had stopped keeping my cards alive

For several months, I'd been using my card box less and less.

It wasn't a decision to give it up. I hadn't concluded that the method had stopped working. I hadn't found anything better. I'd simply been caught up in something else: producing with artificial intelligence.

AI lets you produce at an extraordinary speed. With tools like Claude Code, I can sometimes do four, five or ten times more than before. Every time one task speeds up, another becomes possible. Then another. The time you save isn't released: it goes straight back into more output, more checking, more confirming.

I ended up producing almost without a break.

And I gradually let go of what feeds that output: reading slowly, taking notes, letting an idea resist me, going back over a conviction, producing nothing for a while and accepting that I don't yet know what I think.

The problem isn't that AI stops us thinking. The problem is that it makes producing so easy that you can go on and on without ever stopping to think.

You can hand it three fragments, ask "what do you make of this?", and get a coherent formulation back in seconds. That's useful. I do it. But once the move becomes automatic, AI ends up serving as a cognitive crutch. It steps in before we've framed the problem ourselves, tested the contradictions or found the nuances.

AI multiplies whatever we already are, but it doesn't automatically make us grow.

If I stay intellectually at time N, it can amplify that state with spectacular force. But it doesn't guarantee I'll reach N+1 and N+2. That kind of growth still takes time, curiosity, reading, and the personal work of testing your ideas against something.

That's exactly the time my pillar cards were meant to protect.

A tweet brought an old card back

One weekend, with my computer switched off, I came across a tweet from Olivier Martineau:

"I keep telling students there are two ways to use AI: 1. get the AI to do your work; 2. do better, learn and improve thanks to AI. The difference? If the AI is enough to do the work you've been given, you don't have a job."

That immediately chimed with a card I'd written long before ChatGPT, drawn from The Almanack of Naval Ravikant:

"Specific knowledge is knowledge that you cannot be trained for. If society can train you, it can train someone else and replace you."

A second line completed the idea:

"Specific knowledge cannot be taught, but it can be learned."

In Naval Ravikant's original formulation, the replacement is one human by another human that society can train just as well. The tweet shifted the question sharply: if everything I know how to do can be handed over to an AI, what's left that's specific about my knowledge?

I went and found the card.

Then it called up others: cards on information, knowledge, experience, judgement, creativity, intellectual competence. I took them out of the box and laid them on the table.

I didn't have an answer yet. All I had was a tension strong enough to make me want to think.

The table as a large thinking screen with no interface

I started moving the cards around.

One looked like the consequence of another, so I put it on the right. Another introduced a condition; I slid it underneath. Two cards seemed to reinforce each other. A third set up an opposition. I tried one arrangement, then another.

The layout wasn't decorative, and it wasn't planned. It followed the thinking as it took shape.

The colours mean nothing here. The cards sit close to each other according to the connections explored during the session.

I can link notes in Obsidian. I can build a graph in Miro. Those tools are genuinely useful, but they put an interface between the gesture and the idea. To draw an arrow you have to pick a tool, choose its direction, sometimes set how it looks. To keep the whole picture in view you have to zoom in and out. To move something you first have to select it.

On the table, I pick up a card and I move it. The gesture is enough.

Every card stays visible at once. I can start again from scratch instead of following the connections a piece of software has already recorded. I can try a composition, destroy it in a second and start over somewhere else.

The table becomes a large thinking screen with no interface.

At that point the cards aren't just memory holders. They form a device for composing thought.

Challenge, don't confirm

The word that matters here is challenge.

I don't take my cards out to have them confirm what I've just read. Nor am I trying to mechanically manufacture a bigger thought than the original tweet. I put this new claim up against my pillars: do I agree? Where does my disagreement start? What has changed since the card was written? What nuance is missing?

The challenge can strengthen a card. It can also weaken it, shift it, or show that it no longer holds.

A card doesn't even need to hold an idea I agree with. Some claims shake my certainties hard enough to earn a place in the box. As long as I don't have a counter-argument that genuinely dismisses them, they stay active. A card like that becomes an unsettling pillar.

Leaving them out would mean organising my knowledge so as never again to meet what contradicts me. That would be the exact opposite of the point.

Once the cards are laid out, I take their connections into my notebook. I don't copy them. I draw arrows, dependencies, oppositions and clusters. The page brings something the cards didn't hold separately: a visual synthesis of the thinking produced while they were being challenged.

The notebook page doesn't summarise the cards: it shows how they fit together around the question of specific knowledge in the age of AI.

The Naval Ravikant card wasn't rewritten. The quotation is unchanged. But the network of thought around it has moved. A claim framed before generative AI met a new situation. The notebook keeps that update.

The card carries the pillar. The notebook carries the composition built around it.

For a card to come back, you still have to carry it

That scene depends on earlier work, though — work that's far less visible.

For the tweet to bring back the Naval Ravikant card unprompted, the card has to be present enough in my memory. A quotation stored somewhere in a note base doesn't necessarily produce that resonance. I can find it with a search, but I won't think to search for it unless it has already come back to me.

That's the first use of my pillar cards: keeping a small number of anchor points alive through active recall and spaced repetition.

The Leitner box rests on a simple principle. Cards you have a weak grasp of come back often. The ones that become familiar are reviewed less and less. The repetition spaces out as the memory becomes more durable.

I'm not trying to apply that mechanism to the thousands of atomic notes, reading notes and sources I hold. That would be unworkable. My brain isn't a RAG that can sweep tens of thousands of fragments on demand.

I apply Leitner to a far smaller set: my pillars.

The logic borrows from the Pareto principle. I concentrate the effort of reactivation on a few cards able to call up a wider network. I'm not claiming that memorising a root automatically memorises the whole tree. But while that root stays reachable, it offers a way through to other ideas, other reading and other experiences attached to it.

I currently use three very loose levels:

1. My grasp of the idea is still shaky. The card reminds me of something, but I can't yet unfold its consequences or its ins and outs. I review it roughly every fortnight.
2. My grasp is middling. I can explain the idea and some of what follows from it, without having explored its full depth. I review it roughly every month.
3. My grasp is deep. I can talk about it with some subtlety, challenge it with a real situation, or write a whole article on it. I review it roughly every two months.

These intervals aren't hard rules. I'm not trying to industrialise my memory. Moving from one box to the next isn't a reward for perfect recitation; it expresses how far I've made the idea my own.

What matters most isn't the mechanics of the box. It's the link between the two uses: repetition keeps the pillars available; the challenge stops them setting hard.

A deck that stays small, alive and incomplete

This practice only works while the cards stay rare.

An idea doesn't become a pillar because it's interesting, elegant or useful. It has to produce a much stronger anchor. When I reread it, a few lines have to be enough to bring back immediately whatever it opened up in me.

Otherwise the box ends up reproducing the problem of the note base: hundreds, then thousands of items that exist somewhere but can no longer be kept alive.

I've set myself a personal ceiling of two hundred cards. I currently hold fewer than a hundred, because my entry test is demanding. There's nothing scientific about the number. I took some inspiration from Dunbar's number, often summarised as a limit on how many stable social relationships one person can sustain.

I'm not transposing that finding to the memory of ideas. Dunbar's number is about social relationships, and generalising it into a single human figure has itself been disputed. All it gave me was an intuition: whatever we want to keep actively alive probably has to stay bounded.

That borrowing is itself a good illustration of what the cards do. Knowledge from one field can be moved across, set against a different problem and adapted into a personal hypothesis. I'm not trying to prove that two hundred is the right number. I'm testing a constraint.

I don't yet know what will happen when I hit the ceiling. Will I have to throw out more cards? Replace some? Merge them, at the risk of blunting their individual force? I have no answer.

What I do know is that a card has to be allowed to die.

I have one on the product trio: the product manager, the developer and the designer as the core decision-making team. Given the way I work now, that card has aged badly. With AI, a single developer can take on the scale of a team and come far closer to the business. The old frame no longer seems to carry the same reality.

The active deck isn't an archive of every conviction I've ever held. If it stops representing my current pillars, it loses its function.

Nor should it be filed by subject. The colours of my index cards mean nothing. I don't want an "AI" group, a "product" group or an "innovation" group. A card drawn from Karl Lagerfeld about facing up to ideas can perfectly well meet a card on startups or artificial intelligence.

The connections depend on the question in front of me, not on where the card came from.

Paper composes, digital finds and protects

My system is still paper first.

I use index cards close to A7, big enough for a title and three to five lines, but too small to turn into mini reading notes. For a quotation, I write down the author and the title of the book. I rarely note the page: with digital editions and shifting pagination, a search inside the book finds the passage more easily.

I can write the number of a related card on the back. But I keep those connections light. They flag an association I've already come across; they don't settle every composition to come.

Each card also has a mirror in Obsidian. That lets me search a quotation in full text, copy it without retyping it, and keep a backup if my box is lost or damaged. The digital graph shows the links already established, but I barely use it to think with. The table is precisely what lets me start again from scratch.

When I'm away from home, I can capture an idea temporarily in Apple Notes and copy it onto a card later. The physical card remains the original.

Paper composes, digital finds and protects.

The choice of materials matters surprisingly little. I use Oxford cards and Moleskine notebooks because I want formats I'll plausibly still be able to buy in a few years. Those aren't requirements of the method. What counts is using cards of a single size, sold reliably over time, and constrained enough to keep you brief.

Neither PARA nor Zettelkasten

Pillar cards don't replace my other knowledge management systems.

I still use PARA and Zettelkasten, for different ends. I'll devote a series of articles to them, because both methods deserve better than a few paragraphs of summary.

PARA is strongly geared towards action. A project creates a need, and the available knowledge is then put to work to get a result. Zettelkasten leaves more room for emergence: reading becomes notes, notes connect to each other, and new ideas surface in the network.

Pillar cards sit somewhere else. They're neither a general base nor a project management method. They form a personal, limited core of ideas I want to keep alive inside me.

When I put them on a table, I'm not necessarily asking them to produce an article, a decision or a usable answer. I'm asking them to help me think through a tension that matters right now. Something may come of it, but the session isn't obliged to deliver.

Trying to cash in on that thinking straight away would destroy part of its value. You have to keep the option of drifting, of running into a contradiction, and of finishing with no deliverable but with your thinking somewhere else.

Rebuilding time to think alone

The main risk with this method isn't picking the wrong box. It's never leaving it any time.

Slack, LinkedIn, the daily requests and AI-assisted output easily keep you inside a continuous flow. Inside that flow, curiosity erodes. You no longer read enough for an idea to stop you. You no longer give it time to come back to you. The box then turns into a collection of dead quotations.

So I'm trying to reinstate a regular appointment, ideally an hour a week.

During that hour Slack is off. LinkedIn is off. The computer is off. ChatGPT too.

I'm left alone with the box, the cards, the notebook and whatever I can do with them. I can't immediately ask an AI to interpret my fragments for me. I have to frame, doubt, connect, contradict, and sometimes admit that I don't know yet.

The session can start from a tweet, something I've read, a conversation or a problem that came up during the week. I can set that aside when it comes up, then pick it back up once I finally have the time.

If nothing imposes itself, I can drift instead. I take a few cards at random. I reread one, then another. An opposition shows up. One card calls up a third. The resonance doesn't always come before the session; it can come out of that loose wandering.

This slow phase isn't a rejection of AI, or nostalgia for paper. On the contrary, it prepares me to put AI to work better.

I now separate two phases:

1. slow consolidation, where I read, handle, write and thicken my thinking without AI;
2. fast execution, where I put that thinking to work to drive the tools, produce, analyse or decide.

The speed of the second phase depends on the depth of the first. If I cut out consolidation, I can carry on producing more, but I risk endlessly amplifying only what I already knew.

Start with fifty cards and a pen

The surest way never to start would be to try to install the whole system at once.

Don't buy three boxes. Don't define a taxonomy. Don't pick a colour code. Don't build an Obsidian graph before you've written the first card.

Buy a pack of about fifty index cards and a pen. Pick a simple format, close to A7, that you'll easily find again. Leave the cards visible on your desk.

Then wait until an idea really marks you.

Not just an interesting idea. Not a well-turned sentence you'll have forgotten by tomorrow. An idea that sinks in, that organises something, that lights something up or unsettles you enough that it stays open.

Write a number, a title, then the idea or the quotation in a few lines.

Go back over those first cards from time to time. Ask yourself whether they really bring something back, whether they call up other knowledge, whether they still match your experience. If the practice is giving you something, you can then buy a box, create a digital mirror and gradually bring in spaced repetition.

But start before you've designed the perfect system.

The point isn't to manage all your knowledge. It's to make a few ideas present enough that they can meet each other, challenge each other, and keep changing you.