Second Brain

Every document this firm produces is a plain markdown file in one folder tree. No proprietary format, no database holding the text hostage, nothing that needs a vendor to still exist in ten years. That is not nostalgia. It is the precondition for everything else: an agent can read, write and reason over plain files without an integration, and a human can read them in any editor when the agent is wrong.

A development environment is usually the one thing nobody writes down, which is why rebuilding one takes a week and the knowledge leaves when the person does. This one is documented to the same standard as everything else: installs, runtimes, shell and path, identity and auth, the agent harness, every project checkout, the editor and the fonts, what syncs and what exists in exactly one copy, and a rebuild sequence that starts from a bare machine.

Credentials are in it by name and location only. The section recording them was produced without reading a single value, and says so at the top. Knowing that a key exists, what it is called and which file holds it is what you need to reason about blast radius; the value itself is what you need to leak one.

It refreshes itself. A scheduled job collects the machine's state each week, reads the diff for anything credential-shaped the patterns might have missed, commits and pushes. A check then fails if that backup is more than nine days old or has anything unpushed, and it runs at the end of every working session, so a job that quietly stops firing gets reported rather than discovered later.

The architecture diagrams live in Excalidraw and are referenced from the record rather than pasted into it, so there is one copy of each and no exported image going stale in a folder somewhere.

One of them is not drawn at all. The capability landscape is built: the band geometry, the taxonomy, the placement rules and the edge conventions live in a generator, and the scene is replaced by running it. Hand-editing that one is pointless, because the next build discards it. A diagram produced from the same source as the argument it illustrates cannot quietly disagree with it.

There are structural checks on the output: boxes that overlap, arrows crossing shapes they have no business touching, labels sitting on top of each other, text whose two stored copies disagree. Those are not sufficient and the record says so. Several defects passed every check and were caught only by rendering the thing and looking at it, including a band drawn right to left so the reading order reversed, and a six-row legend that came out as six copies of one label.

The record is verified rather than trusted. A script re-derives every checkable claim the architecture document makes, against the live systems, and fails when reality and the written record disagree. Documentation that is never checked is a story about the past; this is the difference between a diagram and a system.

The standing rule behind it: any statement about a live system carries the command that established it. If no check was run, the answer says so and names what would settle it. That rule applies to the assistant and to the person, and it is the single highest-leverage thing in the whole arrangement.