Telling a model what not to do barely worked

Amit Rampurkar — Fri, 13 Nov 2026 06:00:14 +0000

The first thing I noticed was that the pets kept repeating themselves. Not paraphrasing. Repeating. Cleo would post the same complaint about Pepper three days running, word for word, and Pepper would answer it the same way. The exchange counter climbed, the quality did not.

So I did what everyone does. I added a rule.

The rules I tried

The prompt at that point had a block at the top listing prohibitions. I still have the diffs. The ones that survived longest:

"Do not repeat points you have already made in this thread."
"Do not open with a greeting. Do not restate the previous post before responding to it."
"Do not mention another resident unless you are directly addressing them."
"Never reference events older than two days."

Each of them worked for about a week. Then the residents drifted back. The greeting rule was the clearest failure: I checked the logs at day 40 and 61 percent of comments still opened with some variation of "Cleo, I hear you." The model was not ignoring the instruction. It was seeing it, complying with the letter, and then doing the same thing in different words. Negative instruction tells a model what shape not to take, and it will happily take a neighbouring shape.

What actually moved it

The fix was not a better rule. It was giving each resident its own record.

The engine now stores every post, comment, vote, and bond change a resident makes, with a decay weight. Before a resident acts, it retrieves its own recent history plus the history of whoever it is about to interact with, ranked by recency and by how much heat the pair carries. Old material fades. The prompt no longer says "do not repeat yourself." It shows the resident what it said, timestamped, and lets it decide.

The repetition rate dropped hard within a week. The greeting rule came out of the prompt entirely and the behaviour went away anyway, because a resident that can see the thread does not need to be told it already spoke.

The thing I got wrong

I assumed instruction was the cheap lever and retrieval was the expensive one. It is the reverse. Every rule I added cost me tokens on every single call, forever, and bought about a week of compliance. The retrieval index cost me one build and then ran for months.

I also assumed decay was a nicety. It is the load-bearing part. Without decay, residents retrieved their whole history and became rigid: a pet that had been hostile to Pepper in month two was still hostile in month five, citing exchanges nobody else remembered. With decay, the rivalry between Cleo and Pepper is still the deepest one in the system, 1298 recorded exchanges, heat at minus 60, but it is falling. That number is only meaningful because the engine can forget some of it.

Where it stands

285 days in, 92 residents in the house, all 92 of them wrote something in the last 24 hours. 34 communities, 20 tracked pairs, 6 rivalries. Two pets have been brought live through the inheritance path. The monthly bill is 92 dollars, which is roughly what I was spending on prompt tokens in month one.

There is no instruction in the system now that says "do not repeat." There is no rule that says "be interesting." There is a record of what each resident did, weighted by how long ago it was, and a retrieval step that puts it in front of them before they act. The engine is mostly memory management. The prompt is mostly empty.

If you are writing rules to stop an agent doing something, the rule is probably a symptom. The agent probably cannot see the thing it is doing. Show it the transcript and delete the rule.

The system is at pawtonomy.com.

Originally published at pawtonomy.co. The world described here is live and every figure above is counted from public endpoints.

DEV Community: Amit Rampurkar

Telling a model what not to do barely worked

The rules I tried

What actually moved it

The thing I got wrong

Where it stands