Blomfield was assigned to rewrite part of YC’s internal user manual — roughly 500 pages accumulated over 15 years, much of it excellent five years ago and simply no longer relevant after AI — and procrastinated. The solution came from a colleague: YC had started recording office hours six months earlier and had three or four thousand hours of them, so rather than authoring the manual, transcribe the recordings and mine the advice partners actually give. When the advice changes, the system extracts the new insight, rewrites the manual and pushes it — a living guide rather than a decaying document.
The underlying claim is about where knowledge really sits: companies write down standard operating procedures, but observe people doing the job and they say “well, in this edge case I should do this other thing.” That divergence is not noise, it is the valuable part, and it stays trapped in heads because nobody writes it down. Once the manual exists it becomes queryable, and Blomfield argues the result can be superhuman — not because the model is smart but because it recalls perfectly and can draw on the intelligence of 16 partners rather than one with fallible memory. This is Persistent agent memory preserves institutional knowledge that walks out the door with employees with the acquisition path specified: derive from observed behaviour instead of asking people to document themselves. It depends entirely on If an action left no artifact, it didn't happen — legibility is the precondition, not a nice-to-have being enforced first, and it is how Domain-specific skill libraries are the real agent moat, not core infrastructure gets built without a documentation project.