TL;DR
- 01Polanyi’s observation — we know more than we can tell — names the most valuable and least managed asset in any institution: the structured judgment of its experienced people.
- 02It is also the asset AI cannot reach. Models train on what was published; an HR caseworker’s triage instincts, a mediator’s sequencing, a speechwriter’s forbidden words were never written down.
- 03Elicitation works when it asks for devices, not anecdotes: named procedures, ordered checklists, detection prompts for known traps — typed structure an agent can apply and a reviewer can check.
- 04This is where 1980s expert systems died, and why capsules take a different route: capture as a byproduct of real work plus structured interviews, with the expert reviewing everything that carries their name.
An HR caseworker’s triage instincts, a mediator’s sequencing, a speechwriter’s forbidden words — the most valuable knowledge in any institution was never written down. How structured elicitation turns it into typed, attributable capsule content, and why the 1980s expert-systems failure does not repeat.
Three experts, three invisible structures
- ▸The HR caseworker, 2015–2020. Five years of grievances taught her which combination of timeline gaps and witness posture signals escalation, the interview order that preserves trust, which findings-phrasings survive legal review. None of it is in the case files — it is in how the case files were produced.
- ▸The mediator. Before tabling any proposal he runs an escalation-ladder test he has never written down. He hears positions and extracts interests by habit. He never lets a commitment go unrecorded — a discipline, not a document.
- ▸The speechwriter. The argument arc (concede early, anchor in principle, ask late), the words this audience has heard weaponized, what the principal will never say even when the draft logic calls for it. The speeches exist; the rules that generated them do not.
Why elicitation usually fails — and what works
Ask an expert “how do you do it?” and you get war stories — vivid, true, and useless to a machine. The expert-systems era proved that handing experts a blank knowledge base produces neither knowledge nor a base. Two things changed since: capture can now ride on top of artifacts that already exist (resolved cases, transcripts, exemplar documents — which models are excellent at mining for candidate structure), and the interview can be conversational, with the model proposing typed records the expert corrects rather than authors.
The output discipline matters as much as the input. A device is a named procedure with required inputs, ordered steps, and critique prompts. A trap is a failure mode with a detection question. A contract is a checkable rule about output. These are the units a Tool or Output capsule holds — and the units an agent can actually apply, attributably, in the margin of a draft.
Attribution is the incentive
Tacit knowledge stays tacit partly because sharing it has been all cost and no credit. A capsule changes the deal: the heuristics carry the expert’s name, the review tier records who vouched, and outputs built with the capsule can say whose thinking shaped them. Knowledge transfer with a byline — which is, historically, the only kind that scales.
SOURCES
- [1]Polanyi, M. (1966). The Tacit Dimension.
- [2]California Management Review (2026). Tacit Knowledge Is Your Next Competitive Moat.
- [3]Feigenbaum, E. & McCorduck, P. (1983). The Fifth Generation.The expert-systems lesson capsules are designed not to repeat.
