An AI assistant starts every session knowing nothing about you. You can paste in context each time, or you can write it down once, keep it current, and have the assistant read it before it does anything. I chose the second, and the result is a folder of plain text files I have started calling my operating system.
There is no software in it. One file says who I am and what I know, so the assistant does not explain the basics of fields I worked in for thirty years. One sets the voice: lead with a recommendation, skip the hedging. One lists the tools available and what each is allowed to do. A memory folder holds a dated log of what happened, written one line at a time.
The folder sits inside a directory that syncs between my Macs. That was the first design decision and the one everything else follows from: my work moves between machines constantly, so the context has to move with it. A session on the laptop picks up where the desktop left off, because both read the same files.
Two rules in those files matter more than the rest.
The first is verify before asserting. A language model will state a version number, a price or a statistic with complete confidence and be wrong. So the instructions say: if a fact can be checked, check it before stating it, and if it cannot, say so and label what follows as a guess. For facts about my own world, the web is no help, so the assistant is pointed at my notes as the system of record.
The second is host isolation. Nothing an agent builds may install a runtime on the Mac itself. Everything runs in a container. Agents are productive and careless in equal measure, and after a few months of experiments a machine fills up with interpreters and packages nobody remembers installing. Containers make each experiment disposable.
Writing this down had a side effect I did not expect. Most of it is not about AI. It is the set of standards I would give a capable new colleague: how I want things written, what I consider verified, what you may touch without asking. I had never made them explicit, because a human colleague absorbs them over months. An agent needs them on the first day, every day.
I spent a career writing security policy for organizations. This is the same exercise at the scale of one person: decide what is allowed, write it where it will be read, and check that it is followed.