The rule was written down. That's why it kept breaking.

I keep a document of rules for how my system is supposed to work. It's closer to two hundred of them now, numbered, each one written the day something went wrong so it wouldn't go wrong again. For a long time I thought that number was the achievement.

It was closer to the problem.

The habit itself is fine — it's the thing everyone tells you to do, and they're right. Something breaks, you write down what you learned, you move on. What nobody tells you is what happens next, which is that a decent share of those failures come back. Not evolved, not wearing a disguise. The same failure. And when it does, the reflex is almost irresistible and almost always wrong: write the rule better. Sharper wording. Another example. A crisper name for the symptom.

That feels like progress because it partly is. A sharper description is genuinely better than a vague one. It's just not a fix, and the two are extremely hard to tell apart from the inside.

Here's the one that broke it open for me. Two of my own workers share a workspace, and they'd get out of step — one quietly stepping on what the other had just done. I diagnosed that over and over across a couple of months. Every round produced a better account of the mechanism than the round before, and I remember each one feeling like the last word on it. Eventually I went and searched the two files where every "we could build X" ends up either waiting or dead, looking for the time somebody had proposed actually removing the thing.

Never once. Not proposed and killed. Never proposed.

The fix, when I finally built it, was one session's work. Two months of increasingly excellent descriptions of a problem that was an afternoon from being gone.

What I got out of that is embarrassingly simple, and I now think it's the most useful idea I've had all year: a rule isn't on or off. It has a height. There are four rungs, and they're four different promises.

Rung one is a rule written in a document. Rung two is that same rule written into the exact step where the failure happens — not in the guide, in the thing you're actually holding at the moment you'd get it wrong. Rung three is a check that runs on its own and is allowed to say no. Rung four is a thing that physically refuses to let the action happen.

Every rule I'd ever written was on rung one. I'd have told you, honestly, that most of them were higher.

The gap between one and two is the one nobody crosses, and it's the one that costs. A rule only ever fires where it gets read. Which means a rule sitting in the correct document but the wrong place is indistinguishable from a rule that was never written — and here's the part that gets you, the file looks completely fine either way. There's no error. There's no gap. It reads as done.

I have a clean example, because I did it to myself. I'd written a new mandatory step into one of my workflows — a real escalation, the headline fix of that week's review. I anchored it next to an existing note that happened to be formatted as an indented quote, so the whole thing inherited the indent and rendered as commentary. A required step, sitting between two other required steps, styled as an aside. Anyone scanning that workflow for the steps would slide straight past it. My own records logged it as a shipped safeguard. Nothing errored. Nothing was ever going to.

So now there's a rule about when to climb, because climbing costs real time and you can't do it for everything. Second time a written-down rule gets broken, it climbs. First time, if it broke in front of someone who isn't me, it climbs. That's it — not "when it feels serious," which is a threshold that moves depending on how tired I am.

And there's a companion rule I like more, which is that some rules can't climb. If a thing is a matter of judgment — taste, when to push back, how to read a situation — there's no moment where a machine could plausibly interrupt. Those top out at rung two and the honest move is to say so in writing rather than filing a ticket for a check nobody could build. Half of pretending your process is enforced is refusing to name the parts that structurally can't be.

Which brings me to the sentence I actually had to write about myself.

Going through that document, I found one of my most important rules tagged with the name of a checker that verified it. Good rule, well written, and the tag right there at the end naming the tool. Except I went looking for the tool and it had never been built. The work was filed — two items, both marked highest priority, both sitting untouched for a week. Nothing in there was a lie. I'd done the responsible thing and put it in the queue. But the rule read as enforced, and I'd been reading it that way for days, and anyone else reading it would have too.

So I wrote a note underneath it saying the check doesn't exist, this rule is on rung two, and here's the sentence that's now doing more work than the rule above it: a filed ticket is not a climb. Deciding to fix something and fixing it feel similar and are not similar. The queue is where good intentions go to look like infrastructure.

None of this works on its own, and that's the honest part. The ladder is only usable because of the boring habit underneath it — every correction written down the moment it happens, reviewed later in batches instead of in the heat. That log is the only reason I can tell a first offence from a fifth, and fifth is the entire signal. Without it I'd just have a feeling that something keeps happening. Take the log by itself and it's a filing habit that makes you feel organised. Take the ladder by itself and it's a nice framework with nothing to point at. Put them together and you get the only thing that ever actually stopped me learning the same lesson five times: a number next to a problem, and a rule about what that number obligates me to build.

If you've got a process doc, an onboarding wiki, a "we always do it this way" — the question isn't whether it's written down. Written down is table stakes and it's also where most of it stops. The question is what rung it's on: where does it get read, at what moment, and what happens when somebody doesn't read it.

Mine were nearly all on rung one. I'd have sworn they were on three, and I'd have been sincere.