Monday, August 24. I opened the file that loads at the front of every AI session I run, and I counted.
Three hundred and nine lines. Before I typed a word, the AI had to read all of them.
I started scrolling.
The instruction to be brief was in there four times. Not four slightly different rules. Four copies of the same one. Under them, a lecture on why being brief works. Two rules still came with the whole afternoon it took to learn them. Old decisions, already replaced, sat next to the current ones. The AI read both. Every session.
Somewhere in the middle was a sentence I had already written. Keep this file small.
That sentence was in the file. It had not kept the file small.
It was not the chat
People already have a name for a long chat going stale: context rot. The original instructions are still sitting there, and the AI starts acting like it never read them. That is real. It is also not what I spent Monday on.
The chat was not the problem.
The file that loads first was.
I mean the instructions you write once, so the tool remembers how you like to work. Every tool has a box for this. It loads whether the chat is one message or four hundred. You can start a new conversation. You cannot start a new first page unless you go edit it.
It grows the way a how-we-do-things file grows. Every incident earns a paragraph. Nobody sits down and reads the whole thing as a stranger would.
I did not need more rules. I had been keeping the reasons the rules existed in the same place as the rules.
Saying the same thing four times does not make it stick. It is the same instruction, four times, in the way of everything else, on every message, before the work even starts.
I thought the cut was the work
I moved 235 lines into a file the AI does not open unless someone asks. The live file went to 57.
Nothing was deleted. The lecture and the long afternoons are still there. They just stopped riding along for free.
I thought the cut was the work.
I had a short file. I did not yet have a second pair of eyes. The note I left myself said as much: cleaner was still one opinion.
Then I asked two other AIs, neither of which had done the cutting, to read the short file against the long one.
They both said no.
Not "make it longer." No. The short file had dropped things that were still live. I had mistaken shorter for finished.
The line that had gone blunt
The cut had dropped rules I still needed.
There is a rule about how I want to be spoken to: short, and still a person. Not blunt. The cut had squeezed that into do not pad, match how long I wrote, delete what I can already guess. All of that is still true.
Warmth was the part that fell on the floor.
Without it, be brief reads as do not be considerate. That is not the deal.
Never blunt had to go back as its own line, or the rest of the file would keep teaching the wrong lesson.
A few one-line rules I still use every day had fallen into the archive too. The AI never sees that file unless someone goes looking. Those rules had become stories about rules.
I put them back as single lines. The file is still 57.
Shortening the file by yourself is a quieter way to lose the thing you were trying to save.
Keep this small had already failed once, as a sentence. The live file now has a cap: 100 lines, and a check that refuses the save if it goes over. Next time a paragraph wants in, it becomes one dated line, or it goes to the archive.
What I will do next time
The cap is an alarm. It tells me the file got long. It does not tell me the file is still true.
The first pass is you. The second pass is someone who did not do the cutting. Either one alone will lie to you in a different direction: one keeps too much, the other throws out something still live.
Every four to six months I am going to read that file as a stranger would. Live rules stay, one line each. The stories go in a file the AI does not open on its own. Then a second AI, one that did not do the cutting, reads both, and I put back anything still live. Then I put the next pass on the calendar, before the file starts teaching the lecture again.
Do not wait for the AI to start ignoring you. By then the problem is already in the file it reads first.
Long chats still go stale. Start a new one. That is a different problem, and it does not excuse the file that loads first.
Keep the live rules short. Put the stories where they do not ride along for free.
How long has it been since you read the file your AI reads first? Hit reply. I read everything.
Christopher
Wisdom Nexus is written the way it is lived, with AI in the loop and a person at the desk who cares how it lands. This is where the theory meets the practice, shared with you in the open.
P.S. The scoreboard is still mostly me. The file that loads before I type went from 309 lines to 57 this week, and a second pair of eyes still caught the rules I had dropped. Same class of lesson as the phone case: the copy of the truth is the thing that rots.