The Context We Wrote Down

Episode one of AI-Native Context ended on a gap: ask someone for the SOP and you get fifteen bullet points; watch their screen and they click fifty times. The difference between those two numbers is the context nobody ever wrote down. This week Chris Carolan and Casey Hawkins do the unusual thing and open the file where it did get written down — MainBrain's CLAUDE.md, the operating-rules document every AI agent on the Value-First Team loads before it does any work. Not a summary of how the company runs. The actual file, on screen. What is in it is mostly scar tissue: nearly every rule exists because something went wrong once and nobody wanted it to go wrong twice. "Never smooth someone else's words" is there because fluent writing tidies a quote by default, and a cleaned-up quote reads better than the real one, so nothing downstream can tell. "Declare the population" is there because it is entirely possible to report with confidence on a set you never actually counted. The Value Decrees are deliberately unnumbered, so that citing one by name cannot silently retarget when the list grows. Underneath the specifics sits the question this show keeps circling: what changes in an organization when the context it runs on stops living in people's heads and becomes a file that loads automatically, for every worker, before every task. Every Tuesday.