Four ideas you can take with you

Every project here runs the same cycle: inventory the best of what already exists, distill it into patterns worth keeping, build, rank the results against real taste, and feed what was learned back into the next pass. Nothing about this is novel — it is how careful craft has always worked. What changes with current AI is the cycle time: inventory that took a month takes a day, so the loop spins fast enough to become a practice instead of a project plan.

The full loop needs all three of variation, selection, and retention. Skip variation and it degrades into curation — you only ever choose among what you already know. Skip honest selection and you get output volume with no taste. Skip retention and every iteration starts from zero. Conservative core, wild edges.

Take it with you: before your next project, spend one day inventorying the best existing versions of the thing — the loop only works if it starts there, and with current AI that day actually fits in a day.

“Best” is undefined without a subject. A five-star scale asks people to compress their taste into a number whose meaning drifts by mood, by day, and by person; a head-to-head choice asks only which of these two, for you — and people answer that question reliably. Enough pairwise choices and a preference order crystallizes that no questionnaire could have extracted.

The information-theory framing is literal, not decorative: every comparison removes uncertainty about what someone values. Entropy is the currency. Encode a person’s ordered preferences and you have something a rating never gives you — a taste that can be seen, shared, compared with a friend’s, and consulted by the tools working on that person’s behalf.

That is the thesis this workshop keeps rebuilding in different domains: humans emit comparisons, corrections, and vetoes — not criteria lists and weightings. Systems should listen accordingly.

Take it with you: next time you need someone’s real preference — a client, a team, yourself — don’t ask for a score. Offer two options and ask which. Repeat five times. You will learn more than any survey would have told you.

The working rules, learned the slow way: decisions must live in code where the tooling reads them, not in memory where they fade. Predictions get frozen before results arrive, so hindsight can’t edit them. A null result is reported as a null result. Numbers derive from the authoritative source, never from a convenient intermediate — and a number that can’t state its measurement basis doesn’t ship. The colophon applies these rules to this site itself.

The loop and the ranking thesis converge on one long-term object: a persistent, portable record of what you value — built from your own choices, legible to you, shareable on your terms, and usable by the AI systems working for you. The projects in this workshop are all partial drafts of that object.