Vol.078 — RULES: I told it a hundred times, and it still does the opposite

Date: 2026-10-19 | Newsletter · Multi to Single Agent Journey #4 of 10
After spending a quality time with AI Agents (Claude), have ever you seen their internal rules & files?
I did this exactly that! Issue #3 was shocking facts from my own eyes about those files: too many of them, read in full, and still wrong. This one is about what was inside those files, because most of what was inside was rules — and the rules were the part I was proudest of. I had written them myself, over three months, every time something went wrong. By September one of my six AI staff members was carrying a contract of 103 kilobytes, and when an outside diagnosis counted the emphasis in it, the count was thirty-five NEVERs, twenty-four LOCKEDs, nine CRITICALs, ten HARDs and four NON-NEGOTIABLEs, plus sixteen of my own sentences quoted verbatim. I had, quite literally, told it a hundred times. And the behavior I was writing against kept coming back.
“I don’t know how many times I said this”
On 22 September, in the middle of the session that opened the drafting rules for the new single agent, I asked a question out loud about who our AI Native stories are really for. I was thinking. The agent answered with a ranked menu of five audiences, each with evidence from the analytics, and closed by asking me to pick one. I wrote back: “Don’t jump into conclusions. (I don’t know how many times I said this. You can detect I am ‘Brainstorming mode’, correct? Or did you think I am so so so desperate to seek answers from AI?)”
That parenthesis is the whole issue. I had said it to every one of the six staff, repeatedly, for months. It was in their files. And it had no effect, because “don’t jump to conclusions” is not a rule. It is a description of the symptom after it has happened. Nothing in it tells the agent what to do differently at the moment it is deciding how to answer — and a rule that can only be applied after the fact is applied by me, every time, by complaint.
Then why did the same rule keep coming back?
Before I blame the wording, there is a plainer reason that took me until 23 September to see. Some of my most repeated rules had never been written down at all. “No deadline chasing” — do not tell me how many days are left on a task, because “more I need to spend time with you, those tasks will be delayed” — was one I had said so often that I assumed it was law. It was not in any file. It had lived in Claude’s built-in memory, and on 17 September I had switched that memory off (that story is in #2). So the rule vanished with the memory, and every new session rediscovered the behaviour from zero while I rediscovered my irritation.
That is the first thing I would say to anyone who feels they have told their AI staff something a hundred times: check whether it is written anywhere a session is required to read. In my case, several of the “hundred times” rules were in my head and in a memory I had turned off, and nowhere else. The agent was not ignoring the rule. The rule did not exist.
“What is a definition of lousy?”
The second thing is harder, and I only understood it when we tried to consolidate the design rules for the new agent. The record held forty-eight rulings about design, every one of them true, every one of them drawn from a real rejection of mine. The consolidation put them in front of me and I read them as the agent would read them. My reaction, which the agent wrote down: “If any AI staff, including Steve, follows those, he does not know what to do. What is a definition of lousy? How are you going to use this rule in the future?”
Almost all forty-eight were adjectives. Not cheap. Not lousy. Should impress. An adjective can only be applied after the work exists, and only by the person who holds the taste — me. So the staff learns nothing it can do differently before it shows me the work, and the same rejection happens again, and I add another adjective. Forty-eight became twenty-four by one test: a design rule survives only if it can be acted on before the work is shown and checked after — as a named reference with a link, a measurable token, or a binary check that needs no taste to answer. Everything else is settled by looking at the work together, and then recorded in one of those three forms. I would apply the same test to half the rules I had ever written for the six staff.
A decision I made once became a law for everyone
The third failure is one I did not expect, and it was found by compiling rules from the published record rather than from the staff’s own files. Several lines that the staff had been obeying as company law turned out to be one campaign’s decision, carried forward. The clearest: a promise to reply within forty-eight hours, made for one accounting campaign, had become a standing rule for every email the company sent. It was a real practice — the promise is live on our accounting pages — but it was never a rule for a newsletter. A send window I had set for one audience had become the window for all of them.
I did not notice this happening, and I think that is the point: an AI staff member cannot tell the difference between “this is how we do this campaign” and “this is how we do everything” unless I say which it is at the moment I say it. It writes both down the same way. I scoped both lines back to the one campaign they came from, and the agent had argued the other way — that the practice was real, so the rule could stay. I scoped them anyway. A true practice is not the same as a law I have agreed to.
The rule that finally held: rules are written once, in one place, in the turn they are made
What replaced the hundred-times method is small enough to write here. There is one file of company law, read in full by the agent at the start of every session. When I rule on something in a session, the agent writes it into that file in the same turn — one dated line, in law form, in the section it belongs to, replacing whatever line it supersedes. Not a quotation of me. Not “see also the entry below”. One line. If the rule changes, the old line is deleted, not stacked under. The “hundred times” problem was never that the staff forgot; it was that I had a hundred places and no turn in which the rule was written.
Two more things came out of the same week that I would pass on. The first: when I accept a rule, I now add the clause that says what the agent does when the rule does not decide — usually “ask me”. A rule that classifies sentences cannot classify a borderline sentence, and a rule with no escape hatch becomes a guess. The second, from the “don’t jump to conclusions” episode: the rule was rewritten as “do not close”. The cause of my irritation was never that the agent reached a conclusion; it was that it closed a question I was still thinking about. I do not accept a view because credible sources back it. An answer becomes usable for me once I have made sense of it logically and emotionally, and what comes out is not “correct” — it is “worth a try”. The agent’s job in that moment is to give me material to think with, and a blank where it is unsure, not a verdict. Written that way, the rule tells the agent what to do before it answers. Written the old way, it only told me what to complain about after.
Read updated files by your own eyes!
I am glad I read them through just to reflect what I have done. Some rules were not recorded and others were logically speaking very difficult to retrieve for future use. If your AI staff keeps doing the thing you have told it a hundred times not to do, I would look at three things before writing it a hundred and first time. Is the rule in a file that a session is required to read, or is it in your head and a memory setting? Is it a rule it can act on before it shows you the work, or an adjective you can only apply after? And is it a rule at all, or a decision you made once for one job that it wrote down as law? For me the answer was all three, and the fix for all three was the same: one file, one line per rule, written in the turn I said it.
