The repo that lasted three days.
In May I made a repository and gave it a name so plain it was almost a joke. AI. Three days later I stopped working in it. Twenty-two commits. It is still on the disk.
It was not a failure. It was the first thing I built with these tools that I trusted, and I want to start here because the reason it stopped is the same reason everything after it exists.
Some background. I tried the early chat models in 2023 like a lot of people did. I played for a while, then I pushed one into a subject I actually know, and it gave me confident nonsense. So I left, and I carried the impression out with me. These tools do not really work. They produce false output. I held that view for about three years, and I think a lot of people are still holding it.
What changed it was not an AI story. Coming out of a downturn, the company I work for went from fighting for its life on the floor to a stock price that would not stop climbing, and for months nothing material had changed where I stood. Hiring and capital came later. The market got there first. I wanted to know what it was seeing, and part of the answer was that the tools had moved a long way since I put them down.
One explanation stuck. A system that takes a task in plain English, picks out the tools it needs by their names, and uses them. That is the whole idea, and I have not been able to put it down since. What I wanted from it was specific. Something trustworthy. Something that makes real output I can build the next thing on. Something that checks its own work.
So the first repo. It was a wiki, and I was not the one maintaining it. I put source material in one folder that nothing was allowed to touch. The model wrote and kept the pages in another. Every ingest updated an index and appended a line to a log. Four projects had a page each. It was a good, quiet system and I still think the shape is right.
But it was a chat window with a filing system attached. Nothing in it could go and do anything. I asked, it answered, I filed. What I was learning was not how to build with these tools. I was learning how far I could trust one, and at what level. The answer was: at the level of the sentence I typed.
That is where the one thing it taught me comes from. On the third day I wrote a note to myself, and the system filed it as the only synthesis it ever made.
I have to ask the model to do the right thing. If I can’t ask the right questions, we don’t go down the right road.
It sounds small. It has turned out to be the rule under every rule I have written since. The model goes down the road the request implies. It does not go down the road I meant. When those are the same road it feels like magic. When they are not, nothing rings a bell, and the work comes back polished and wrong. So the question is the leverage. Not the tool, not the model. The question. And getting better at this is mostly getting better at saying what I actually want, in a form that can be acted on, before the acting starts.
Three days was enough to learn that and not enough to use it. I wanted the system to run looser. I wanted it to go and do things without me in the middle of every step. And I had no idea yet what that would cost, or how much of the cost would be rules I wrote myself.
That is the next repository. It lasted four months and two thousand commits, and somewhere in the middle of it I wrote down that I was a little afraid of what I had made.
Part two is the next piece.