Method

I develop almost everything presented on this site with AI agents, in Claude Code. This page says how: what they are given, what is checked and how, and what is never left to them. It describes my practice as of 28 October 2026; it will age, and I date every update.

The architecture decides, the code follows

On a complex project, an agent left to itself solves the problem at hand and undoes the structure: one local fix after another, the project ends up as a plate of spaghetti. BPscript, FaustScript and Kanopi are too complex for that. So they go through Roomi’s Grillhouse, a framework where every decision is written into the architecture before it is coded.

Before any code, a session questions me, turn by turn: goals, reference products, consumers, axes of change, components and interfaces. Then every ticket follows the same path. One agent writes the plan, another writes the tests, a third writes the code that makes them pass, without being able to touch the tests; a fresh agent reviews the whole before integration. The main branch stays green: every integration requires the full check.

Context rather than a bigger model

An agent that searches badly invents: it misquotes what it did not find. I saw it on a corpus of about 2,000 regulatory texts, and switching models changed nothing. Since then, every project is indexed with RTFM, which gives the agent the right passage rather than the whole repository.

A quotation is checked word for word

A quotation written from memory and sourced afterwards ends up wrong. paper-trail enforces the reverse order: the source first, the text second. A reference goes in only if it exists, if the document obtained is the right one, and if the quotation is found in it word for word. What cannot be checked is not quoted.

What is written on this site

The texts follow writing charters, one per genre: course, design, research, method, notebook, project page, reference. They set the intended reader, the outline, the expected evidence, and a list of writing tics to remove.

When an agent played a part in a text, a “How it was made” note says so at the foot of the page: the tool, what the agent did, what was checked. I review everything that is published, and I leave no sentence in it that I could not explain.

What is never given to an agent

Decisions. What a rule settles, the pipeline settles; the rest comes up to me, with the context, an example that shows the problem, the question and a recommendation. I decide, and the decision is written into the architecture or the charters.

My interests

I also advise organisations on AI, with noos-ia. I have no tie to any model vendor.