The method is ready. The instrument is being built in the open.
The old problem was not that experts had no knowledge to give. The problem was that their most valuable knowledge had become automatic. It lived in judgment, timing, sequence, exception handling, and feel. Ordinary interviews could not reliably reach it.
Now there is a way to reach it.
96 turns expert knowledge into structured, inspectable, reusable information. From a 37-year method of knowledge elicitation, it was built for AI to help subject matter experts surface what they know before it disappears.
For decades, most organizations experienced knowledge transfer as mysterious, expensive, and unreliable. So they waited. They documented procedures. They ran exit interviews. They asked experts to "write down what they know."
That was never enough.
The expert knowledge that matters most is rarely sitting in a clean paragraph. It appears when the line should be shut down before the metric says so. It appears when the veteran knows who to call, what to ignore, what sequence to follow, and which small anomaly changes the whole situation.
Not because experts are holding back. The knowledge is there; it has simply stopped coming when called.
Ask directly and you get the compressed version, the shortcut the expert now runs on, because that is the only version introspection can reach. The full reasoning still exists. But it is stored the way a reflex is: retrievable by trigger, not by request. It surfaces when something specific reaches for it (a real case, a real constraint, the right question), not when the expert is asked to summarize.
96 reverses that compression.
It asks structured questions across the kinds of knowledge an expert actually uses: facts, procedures, timing, people, body sense, awareness, systems, and story. The questions do not merely collect answers. They decompose expertise into information that can be inspected, tagged, transferred, and reused.
Your written procedures are not the whole operation.
The real operation also depends on the person who hears the machine differently. The manager who knows which vendor will actually answer. The field expert who sees a pattern before the dashboard does. The senior operator who knows which rule is safe to bend and which one is not.
The difficulty is that flat questions do not reach it.
"What do you know?"
"What should we document?"
"What should the next person understand?"
These questions are too broad. They ask the expert to perform the structure themselves. 96 supplies the structure. It asks the right kind of question for the right kind of knowledge, then turns the answer into something the organization can actually use.
The retirement clock is real. The onboarding clock is real. The knowledge loss is real.
The only question left is whether you do it before the knowledge walks out.
Knowledge Harvesting has three chapters.
Three decades of doing this by hand, one expert at a time. A structured-questioning method with a paper trail older than most of the field.
The AI version was built in public, dated, and tested in the open. The work left footprints while it walked. I have also written dozens of LinkedIn posts that assert the trajectory, share the learnings, and announce the milestones as the method became AI-enabled. See the public LinkedIn activity trail.
96 becomes the final instance of the engine: the same elicitation logic, now being built to run through AI so the method can scale without requiring the original practitioner in the room.
It began with a master's thesis at Auburn (EOR: Establish, Organize, Represent), directed by Charles A. Snyder and written under the influence of Doug Engelbart's unfinished work on augmenting human intellect.
The thesis included Tutor, a structured-questioning subsystem and the direct ancestor of the 96 questions.
LearnerFirst followed in 1992. Then came decades of engagements, field use, revision, and pattern recognition.
The point is simple: 96 did not appear because AI made prompting fashionable. AI made it possible to scale a method that had already been worked out.
When frontier models arrived, the engine already fit. The work was built in the open and dated as it went.
A human expert and an AI model are not the same kind of thing. But both hold usable patterns in language. Structured questions are how you surface what flat prompts miss.
The public trail includes:
This timeline is not a reading assignment. It is the trail.
Humans and AI do not hold knowledge the same way.
Pointed at a human, the engine surfaces knowledge that existed and was hidden.
Pointed at a model, it surfaces patterns: the residue of human cognition in training, language, and use.
Humans hold knowledge. AI holds patterns. Information is what moves between them.
That distinction matters. We keep it visible.
You experience it as a system prompt. It feels simple because the complexity has been absorbed into the instrument.
Behind the surface are eight knowledge types, a 96-cell grid, an orchestration logic, and a decomposition pipeline that turns a live answer into structured information.
The product is not "a prompt." The prompt is the delivery surface. The product is the calibrated question set, the sequencing, the grammar, and the judgment about which question belongs where.
You are sold the surface. The architecture is why the surface can be trusted.
We tested the obvious threat.
A competing frontier model was given a generic description of 96 and asked to rebuild it.
It could derive the outer architecture: intake, parsing, classification, matrix, workflow.
Then it was asked for the thing that makes the architecture work: the empirical calibration. Which 96 questions? Why those? In what order? For which kind of knowledge? Under what operating conditions?
That is where the rebuild broke.
The architecture is derivable. The calibration is not. The text of a prompt can be copied in a keystroke. Thirty-seven years of knowing which question to ask, in which order, for which kind of knowledge, cannot be lifted that way.
That is what you are buying.
A master practitioner's elicitation used to require the master in the room.
That made knowledge harvesting powerful, but hard to scale. The engagement ended when the practitioner left. The cost stayed high. The knowledge often remained trapped inside the method.
96 changes the economics.
It lowers the marginal cost of structured elicitation while preserving the fidelity of the practitioner's judgment.
The claim is not "we scale." Everything scales now.
The claim is that we scaled the thing that usually does not scale: expert-grade elicitation.
Behind 96 sits the larger engine.
It has appeared four times: Tutor (1989) → Builder-Player (the patent) → the Knowledge Harvesting steps → the 96 cells.
Same engine. Different surfaces.
The engine makes implicit knowledge explicit through structured questions.
Its honest limit is part of its strength. The mechanism is general, but the output is not the same in every case, and we do not pretend otherwise.
Doug Engelbart's vision was never only the mouse or the window. Those were the artifacts that got commercialized.
The deeper project was the co-evolution of tools, methods, training, and human capability. He called the goal collective IQ: improving how well people, together, solve harder problems faster.
Knowledge Harvesting is a 37-year contribution to that unfinished agenda.
The connection to Engelbart is not decorative. It is load-bearing.
The work was always about augmenting human intellect. AI did not replace that agenda. AI made it urgent.