Working with AI as a practice

I split thinking from building. One conversation holds the concepts; separate coding sessions build each prototype. They never talk to each other directly. They talk to a folder — short, dated briefings written as prose, a canon of shared definitions, and an append-only log.

Two observations drove the method. Context does not get better as it gets fuller. And decisions need their reasoning attached, because the reasoning is what the next session needs — not just the outcome.

The coaches started talking

An unexpected effect: at some point the prototypes began addressing each other. One coach proposed a change to the shared navigation; the other two picked it up and shipped it. I didn’t mediate. I forwarded a file.

Architecture that travels

What we built doesn’t stay on one laptop. I packaged the governed context as skills and a connector, so colleagues can use approved knowledge, voice contracts and draft checks inside an ordinary AI conversation — no new interface, no setup. Product videos were made from the real interface only, with a written list of what was live and what was illustrative.

The prototype is the pitch

The same thinking became a proposal for how product teams could work now that AI makes building cheap: from the factory, through the workshop, to a studio — where the working prototype is the pitch, and the scarce resource is judgement.

This site is built the same way: plain files, readable instructions for whichever AI edits them, and a human who approves every change.