Fog Show

Never ask the user to verbalize taste — "what does modern mean to you?" produces noise. Show, and read the reaction.

One decision per round: one axis a reaction can be attributed to, every other property held constant. Produce 3–5 variations that are wildly different, not shades — if you can't name a plausible user who would react differently to two of them, replace one. Note the belief each bets on ("this one bets density beats whitespace") in the filename or your notes — never inside the artifact, so the reaction is to the artifact, not the thesis.

Cheap and disposable: visual → one self-contained HTML file, fake data, no wiring; approaches → a one-screen sketch each (the idea, what it optimizes for, its sharpest tradeoff).

The deliverable is a sentence, not the artifact: "you consistently rejected X, so the real requirement is Y." Write it to the ledger as resolved — it is now spec — and delete the probes; the sentence is the only survivor. Reactions that reveal new unknowns become Open items with resolvers.

If nothing lands, the decision space was framed wrong: derive the new axis from the rejection reasons — when every rejection complains about the same non-axis property, that property is the real decision — and rerun.