Elicit, don't just ingest

The pitch for an AI-powered second brain is always the same: feed it your documents, let them link, ask the pile questions. What no version of the pitch tells you is how to teach it what you know but have never written down.

That omission is the whole problem. Expert performance runs on tacit knowledge — cues you read without being able to name them, pattern-recognition that fires before reasoning starts. A pile of ingested documents holds the explicit residue of that: conclusions, artefacts, whatever happened to get typed. It does not hold the perception underneath. A second brain built only by ingestion cannot do what its owner does, and no amount of further ingestion fixes it, because the missing knowledge was never in a document.

But between the documents you happened to write and the knowledge that is irreducibly tacit, there's a wide band you couldstate — if someone asked the right way. Nobody asks. The method for asking exists and has for decades: Cognitive Task Analysis, and specifically the Critical Decision Method, walks an expert back through a hard case, probing at each turn for the cues they read and the options they discarded. That captures the decision context — why a call was made, against which signals. It's the most a repository can hold of expert judgement, and it's exactly what document-ingestion leaves on the floor.

So the move is: don't just ingest, elicit. Interview the owner the way a cognitive task analyst interviews an expert, and file the results as first-class material. I run my own knowledge system this way.

One honest caveat, stated against myself: elicitation raises the floor of what's captured and the owner's confidence in the capture at the same time — and the confidence rises faster. Even a well-elicited second brain is a lossy compression of what its owner knows. Treat it as one.