Small models, long notes
What changed when I stopped asking a model to know things and started asking it to read mine.
For a while I used language models the way everyone did at first: as an oracle. Ask a question, get an answer, check it somewhere else because it might be invented.
What changed my mind was turning it around. Instead of asking a model what it knows, I started giving it what I know, in the form of my own notes, and asking it to read them.
Notes are the context
I have kept notes for years. Meeting notes, half-formed ideas, the reasons we chose one design over another. None of it is organised well enough to search by hand. But a model does not need it organised. It needs it present.
When I ask “what did we decide about retries, and why”, the useful answer is not a general essay about retries. It is the paragraph I wrote in March, found and quoted back to me.
Smaller is fine
For this kind of work a small, fast model is enough. It is not reasoning from the world; it is reading from a page I gave it. The quality of the answer depends far more on the quality of the notes than on the size of the model.
Which means the best investment I made in using AI well was not a subscription. It was writing things down.