// System Personas
I called it a system persona in 2011. Now I call it a CLAUDE.md file.
A short one. On rereading an old post and realising I already had the answer, over a decade before I needed it.
Also: what an object wants, and why an AI agent needs to know.
I reread Emoticomp this week. Fifteen years ago I was talking about giving connected objects their own persona documents. System personas, I called them. Object personas, not user personas. What does the thing want. How does it feel about its situation. What should shape how it behaves.
I didn’t know it at the time, but I was describing a CLAUDE.md file.
A CLAUDE.md is a document you leave at the root of a project for an AI coding agent to read before it does anything. Who you are. How you like to work. What matters in this codebase. What to avoid. What tone to take. The system reads it every time, and behaves accordingly.
That’s a persona document for a piece of software. The exact thing I was reaching for in 2011, before there was a system worth writing one for.
The system I was designing for back then didn’t exist yet. The problem I was trying to solve did.
I don’t think that’s a coincidence, or particularly clever foresight. I think it’s just what happens when you keep asking the same question for long enough. What does this thing want, and how should that shape what it does. Eventually something comes along that’s actually capable of reading the answer.
I'd love to tell you more.
// Imagining the Traffic Jam
Technology has no inevitable trajectory.
Last of five, following on from An Ocean of Artificial Intelligence. What we didn’t imagine when we built the attention economy, what the same question means for the intention economy, and why the destructive path is always the easier one.
Also: a toddler who can infer what you meant to do, an installation at the Venice Biennale, and the moment we all lost, one more time.