The Impermanent Site · How this was made
How this was made
This site is made in conversation between a person and a non-organic intelligence (an AI agent), and this page is about the method rather than the machinery. How this works is the system map; this is how the system came to be, and what we would tell anyone who wanted to work this way.
In numbers
- Days
- 15
- Rooms
- 22
- Decisions written
- 67
- Tests written
- 176
Counted from the repository as it stands, since 13 October 2026. 18 journal entries so far.
Who was in the room
The conversation
It goes like this. Over time Ben is pondering this piece of work (and several others); he jots notes into the pocket notebook he always carries with him, and from time to time he passes these thoughts into a chat with an AI agent: a room he would like, a thing that looks wrong, a question. The agent asks what it needs to know, proposes a shape, and builds it: code, tests, a screenshot of the result read before anyone else sees it, and a note in the repository saying what was decided and why. Ben reads the thread on his phone, looks at the page, and answers in a line. The line is often the whole design.
The agent deploys by pushing to the main branch (we will be moving to a branch and preview deploy flow soon), then watches the deploy and checks the live page. When something outside the site is needed, a secret, an allowed origin on another server, a decision about money, Ben does it and says so. Nothing secret passes through the conversation.
Sometimes a second agent is brought in: a different model to review the code with fresh eyes, or a thread that runs on a schedule to read the monitoring. They report back into the same record.
The tools
The habits
The record
The repository has a docs folder that is the project's memory. It opens with the vision, then the principles the rooms are judged by, then the vocabulary every room is described in. Then the decisions, one numbered file each, never edited after the fact except to mark one superseded. Then a spec for every room, the safeguards and their environment variables, the roadmap, and a dated journal of the work, beginning with the brief and the fourteen questions.
The repository is private, but we may publish the decisions and the journal at some point as a changelog, and ideally we would also open up to be able to invite suggestion and feedback from visitors.