The CLAUDE/ folder
November 2025 — March 2026Prologue — The CLAUDE/ folder
Verified sources: commit
623a7d3(SuperOAuth, November 18, 2025) · commitb0bb88a(March 12, 2026, 02:09)November 2025 — March 2026
In November 2025, there was a CLAUDE/ folder inside SuperOAuth. The idea: give the AI its context once and for all, instead of re-explaining the architecture at every session. 3,255 lines of markdown from the very first commit. Careful. Well structured.
This folder did not come out in one go. It came from several days of back-and-forth with the AI, never committed — including sessions that stopped because the tokens ran out. I kept asking myself « pourquoi je crame mes tokens comme ça » (“why am I burning through my tokens like this”). I wasn’t sure what I was doing made sense. But I « sentais que c’était important » (“felt it was important”): looking for a way to give the AI its context, I was discovering the architecture behind it — without really knowing what I was building.
The problem: it didn’t really work.
Not because it was badly written — it was well written. The problem is that a Claude session starts from zero. You load the guide, Claude reads it, but there is no persistence of reasoning, no memory of past decisions, no context on “why we did this rather than that”. Every session, I re-explained the architectural choices. Every session, Claude forgot the bugs we had debugged the day before.
On November 19, the folder is pruned: 1,043 lines removed, 506 added (b32bd9d). Then nothing. Between November 20 and March 11, SuperOAuth moves forward by nine commits; the guide, by none. Static documentation does not follow the project it describes.
The mental shift: static documentation and context memory are two different problems. The first can be solved with well-organized markdown. The second cannot — it takes an architecture.
What if we applied the same logic as git — where every decision is versioned, traceable, reversible — to knowledge itself? Not documentation files, but a system where every decision is a commit, every agent is a micro-service tuned to that corpus, and context survives session reboots.
The brain began the day I stopped asking “how do I write a better prompt” and started asking “how do I build the coordination layer on top of git”.
March 12, 2026, at 2:09 in the morning. Commit b0bb88a.
Ten days later, on March 22, CLAUDE/ disappears from SuperOAuth. The commit that deletes it already calls it a « proto-brain » (26a046c).