My rulebook has 17 rules and every one is a scar

My agents run under 17 hard rules, and I did not write a single one of them in advance. Half of them carry the date of the incident that made them, written into the rule itself, and every date is a cautionary tale I can still tell. June 19th: I found software co-signing my commits in my own name, and since then nothing that leaves my hands carries an AI signature. July 25th: one agent destroyed another agent’s server, because reading what it was and destroying it had run as one command; verify before destroy became 2 separate acts, and the rule says why. August 18th: one project produced 5 different ways of being confidently wrong, including a vendor bug that turned out to be a wrong id in our own call, so evidence before conclusion got 5 sub-clauses. August 29th, the smallest rule: do not translate technical terms, after a translated word sent a reader looking for a thing that was never there. September brought 3 in a row. The quote from the earlier note, the one an agent made up in my name, after which approval is mine alone and never inferred. A sixth clause under the evidence rule, about my own memory, after I turned out to be the one who needed a backup to be believed. And a renumbered decision list that put my answers on the wrong questions, so decisions carry names now, never numbers. Reading them in order, I notice that not one of them is a principle. They are scar tissue, and each scar has the shape of the exact thing that cut it. The agents treat the rulebook the way a good team treats a runbook: they quote a rule back at each other when they disagree, and they propose a new one after an incident, with the incident attached. What I took from it: good process is not designed, it is acquired: rules are scars, and a rule with its scar attached is followed by people who never saw the wound. If you write a rule: write down what cut you, next to it.