Brain
Running
The system behind every project here: one folder of plain-text files that every AI tool I use reads first. My facts, my rules, my decisions, in one place.
- 3AI platforms, one rulebook
- 9projects built on it
- 10 / 10drift-probe checks passing
- 25,000+model calls logged
HOW IT WORKS
Architecture
How a request moves through Brain: who gives the input, which engine does the work, where the rules and memory live, and what comes out.
Interactive diagram. Select a card to highlight what it feeds and what feeds it; press Escape to clear.
UNDER THE HOOD
What it actually does
One rulebook, every tool
Claude, Antigravity and my local models each start from a small loader file that points to the same rules. Nothing is copied, so nothing drifts.
Found by testing, not assumed
One tool turned out to load nothing at start-up. That test is why every tool now has its own loader.
A drift probe on every change
Ten questions aimed at known failure modes, like inventing dates or flattering me, re-run whenever a rule file or a model version changes.
Routed by the cost of being wrong
Each task is matched to an engine by one question: if this output is wrong, what catches it?
Every output names its engine
Files a model makes record which engine made them. An output with no engine recorded is not treated as evidence.
Numbers checked by code
Any number a local model pulls from a document must appear in that document. Code checks it, not judgement.
Measured before trusted
Local models are tested on real material before they get a job, and replaced when they fall short.
Three agents, one vault, no collisions
Agents re-read a file before writing, edit in place, and never rewrite a decision log.
A list of what I chose not to build
Scrapers that break platform terms and tricks to game AI screening were cut, with the reasons written down.
The nine projects
HOW I DESIGNED IT TO WORK
The rules underneath it
Three lines from the rules I wrote for Brain. Every project inherits them.
“The file before the skill, the skill before the connector.”
On build order
“Work is routed by the cost of being wrong, not by how hard it looks.”
On choosing an engine
“Point, never restate.”
On keeping one source of truth
WHAT'S NEXT
Next, in order
Prove the vault works from a second machine, not only this Mac.
Close the one rule drift the probe caught, so every tool behaves the same way.
Decide how much a cloud session may hand to my Mac on its own.