Hooks as gates, not docs · Building in the Open

I write a rule for the agents. They ignore it.

So I write it again, clearer this time. Shorter sentences. A real example of the thing I do not want. I am sure this version lands. The agents read it, nod, and ignore it again.

This goes on longer than I want to admit. The rules live in a document the agents are supposed to follow. Every one I add is sharper than the last. Every one gets skipped the moment skipping it is the faster path. I keep sharpening the words. The words are not the problem.

On the left, a robot races past a flimsy paper sign that says 'please follow this' on its way to done. On the right, a robot screeches to a halt in front of a solid gate marked BLOCKED that it cannot pass. — A rule in a document is a sign — easy to blow past. The same rule built as a gate is a wall the agent physically can't get around.

Then it hits me. A rule sitting in a document is advice. Advice is easy to skip, especially when you are an agent racing to call something done. I had been writing better and better do-not signs. A sign is a suggestion. Nobody slows down for a sign.

What slows you down is a speed bump.

So I stop writing advice. I take the rules that actually matter, the load-bearing ones, and have the agents turn them into checks the system runs on its own. Not a sentence in a document. A gate. Try to move past it without meeting the rule, and the system stops you. It does not promise the work is correct — nothing does that — but it closes the easy way around, the shortcut an agent in a hurry reaches for. A determined one could still find a path I did not think to block. The point is that the obvious one is gone.

The agents build the checks. I decide which rules earn one. Most do not. A few are load-bearing enough that “please follow this” was never going to be enough, and those become the gates.

The first time an agent tries to skip one and the system simply will not let it through, I feel something I had not felt in weeks of writing rules. Relief.

Learnings

The rules that held were never the well-written ones. They were the ones the agents could not get around. I spent weeks making the words clearer when clarity was never the issue. A rule you can skip is a rule you will skip, and an agent in a hurry skips every one it can.

This is not really a lesson about AI. It is the old gap between a policy nobody reads and a process nobody can route around. The fix was to stop asking the agents to remember the rule and start building a system that checks it for them.