The Lego Box Trick

If you’ve heard people say “ICM” and you still don’t really know what it means, this is another way in.

Start with a box of Lego. Two thousand pieces, no order — you wouldn’t know where to start. Lego fixes that with a sheet: step one, step two, sometimes even separate little bags so you’re not digging through everything for one piece. One bag, one job. A pile becomes a castle, one step at a time.

That’s the idea. Now let’s actually build one, so you can see how the pieces connect — not just what they’re called.

Say your job is writing one newsletter, every week. Here’s what that looks like, built the ICM way.

There’s one small note, sitting right on top of everything, that you’d read first. It doesn’t hold your rules, and it doesn’t hold this week’s draft. It just points: “Rules are over there. This week’s folders are over here.” Some people call this file `CLAUDE.md`, or `AGENTS.md`. The name doesn’t matter. The job does — ten seconds to read, so a person, or a helper, always knows where to look first.

There’s a folder called `_rules`. See that little mark in front of the name? It’s on purpose — it means “this folder is about how you work, not what you’re working on this week.” Your voice. Your style. The way you always want a newsletter to sound. Written down once, so nobody has to explain it out loud every single week.

Then there are folders for the actual steps, numbered so the order can’t get lost:

`01-find-an-idea` — its only job is to end with one sentence: what this week’s newsletter is actually about. That sentence is all it leaves behind for the next folder.

`02-write-it` — it doesn’t open last month’s newsletters, and it doesn’t need to re-read your whole history. It only needs the one sentence `01` left behind, plus your `_rules` folder for how you sound. It writes the draft. When it’s done, it leaves the draft behind.

`03-check-it` — it opens `02`'s draft. Not `02`'s whole folder, just what got left behind. It reads it once, all the way through, and either says yes, or sends it back with a note on what to fix. Nothing moves forward until a person has actually looked at it.

`04-send-it` — it only runs once `03` says yes. It doesn’t rewrite anything. It just sends what’s already been checked.

Here’s the part that’s easy to miss: each folder only ever opens what the folder before it left behind, plus `_rules` if it needs to sound like you. Folder `04` never has to read folder `01`'s notes. Folder `02` never has to re-read every newsletter you’ve ever written. Nobody re-reads everything from the start, every single week.

That’s the actual trick behind ICM. Not the folders by themselves — the fact that each one only has to know about the one folder before it. That’s why it’s called a structure and not just a pile of labeled boxes: the folders are wired together, in order, each one handing off exactly one thing to the next, with a person checking before anything moves on.

And it’s why a helper — an AI, a new hire, future-you in six months — can walk in cold and pick up the work. They don’t need the whole history explained out loud. They open the one folder for the one job in front of them, read its note, and know exactly what to do and what happens after. The structure carries what your memory used to have to.

None of this is something I made up on the spot. A builder named Jake Van Clief wrote it down properly, with a partner, David McDermott, and gave it a name: ICM. If you want the full argument behind it, it’s a real paper — Interpretable Context Methodology (Van Clief & McDermott, arXiv:2603.16021). Everything above is that idea, in the plainest words I could find for it.

So — if you built this for your own weekly thing, what would `01` actually be? Not the whole system. Just the first folder.