What to remember from this issue
Anything you tell an AI in a chat is gone tomorrow. Write your standing instructions in one file, keep its versions, and every AI you use from now on starts from it.
The field note
This week one of my own rules caught me. The failure is more useful than the rule, so here it is.
I keep a written contract for my AI agents: one file they read before every session. Who we are, what our words mean, which rules never bend. I had just rewritten it, a big revision with a glossary. I was pleased with it. I never saved it into version control, the system that keeps every past version of a file.
A few hours later, a second AI session working in the same project overwrote the file. It did not know my revision was there. It was doing exactly what it was asked. My work existed in one editor window and nowhere else, so it stopped existing.
The sentence I wrote after rebuilding it: an uncommitted spec is just a prompt. In plain words, a rule that is not saved somewhere with a history is a chat message. Used once, then gone.
That is the whole problem with chat windows. The AI that starts tomorrow knows nothing yours learned today. So it feels like making the same corrections forever: a brilliant workforce with no memory, where nothing you say adds up.
The fix is not better prompting. The fix is a written way of working: one file that says how work gets done, checks that catch mistakes, and the habit of turning every failure into a permanent rule. Prompts are spent. Written rules accumulate. But only if they are saved with a history, as I was just reminded.
Now the confession. I built a software factory, three times: a system for running fleets of AI agents, with logs, rules for merging their work, the lot. The third version works. I do not use it. Once that one page existed, the one that says how work becomes done, the machinery was unnecessary. One strong model reads the page and hands tasks to smaller ones. A text file replaced a platform.
I see the same move at every scale. A well-known solo developer runs his whole fleet of machines on text files. Palantir sells it to the Fortune 100 under the name “ontology”. The difference between my files and their platform is budget, not kind.
The written description of how your work is done, in words you own, is the only part of the AI stack that gains value over time. Models get better and cheaper on someone else’s schedule. Your description is yours, and it drives whichever model wins next year.
Try this week
Thirty minutes. Take the instructions you keep retyping to your AI and put them in one file. Put that file somewhere that keeps old versions: a shared drive with version history, or a git repository if you have one. That file is the start of your written way of working, and the smallest one that exists.