I've been building something I've been calling an AIOS. A personal AI operating system.
The idea isn't that complicated. Instead of starting every AI session from scratch, you give the model structured context about who you are, what you're building, what you've already decided, and what keeps slowing you down. You store that in files. The agent reads them before it does anything.
It sounds like a simple quality-of-life improvement. It turned out to be something I hadn't fully thought through.
What "Giving Context" Actually Means
Most people use AI tools the same way: they open a chat, describe a problem from scratch, get a response, close the tab, and repeat the whole thing next time.
That works. But it means the AI never learns what you actually care about. It doesn't know you tried a similar approach six weeks ago and it didn't work. It doesn't know that you're 20, building this thing mostly alone, and that the main thing blocking you isn't skill. It's follow-through.
The context layer is an attempt to fix that.
What I ended up with was a folder of files. An intake doc covering my goals and constraints. A connections file tracking every system I've built or wired together. A decisions log where I'm supposed to record what I chose and why. A voice file that explains how I write. An audience file. A frameworks folder.
And a top-level instruction file that tells the agent to read all of it before starting anything.
The Part That Surprised Me
I expected the productivity improvement.
What I didn't expect was how strange it felt to actually write those files.
The intake questionnaire has questions like: what are your 90-day goals? What's the main thing blocking you? What does your business do right now, and what do you want it to do? What do you keep putting off?
I've thought about versions of these questions before. But I hadn't answered them clearly, in writing, somewhere a tool would actually use.
Writing them down and then watching the agent read them and respond to them specifically, in a way that referenced what I had actually said, was disorienting in a quiet way.
It's not that it was magic. It's more that it removed the usual vagueness. The AI couldn't just give me general advice anymore. It had to work with what I had actually said my situation was.
That turned out to be more clarifying for me than it was for the agent.
What the Decisions Log Exposed
The decisions log is supposed to be simple. You make a choice, you write down why, you move on.
I kept not doing it.
For weeks, I'd make decisions about which tools to use, which projects to pause, how to structure the client work I was trying to start, and I'd just carry them in my head. The log sat there mostly empty.
When I finally went back and tried to reconstruct what I'd decided and why, I realized I couldn't for most of it. I'd built things in ways I couldn't fully explain. I'd chosen tools without a clear reason. I'd pivoted on things and couldn't remember what I was pivoting away from.
The log wasn't failing because I was too busy. It was failing because I wasn't actually deciding. I was just drifting and calling it progress.
That's not a tool problem. That's a thinking problem. But the tool made it visible.
The Honest Part
I'm not going to pretend I'm using all of this perfectly.
The decisions log is inconsistent. The connections file has gaps. I haven't filled out the intake questionnaire as honestly as I probably should.
Part of what I'm noticing is that building a good context layer for an AI agent requires the same thing that's hard about any personal system: you have to actually know yourself clearly enough to describe yourself accurately. And most of the time I don't.
I know what I'm working on. I'm less clear on what I actually want, what I'm avoiding, and what's slowing me down in ways I'm not willing to fully admit.
The AIOS setup keeps bumping into that gap.
Where I'm At
I think the setup is working, in the sense that the AI sessions I have now are more useful than they were before. The agent knows my goals. It pushes back sometimes. It references things I've said before. That's real.
But what I keep coming back to is this: the context layer is only as good as the clarity I put into it. An AI that knows vague things about me is only a little better than one that knows nothing.
The part that's actually hard isn't the tooling. It's the honesty.
I'm still figuring out how to do that right.