Taste as Infrastructure

The biggest jump in quality we ever got came from doing the obvious thing backwards.

The first version of the content system we build did what instinct says to do. It read a brand's unstructured material, all the decks and guidelines and back catalog, codified it into a structured set of rules, and then generated new work from the applicable subset of those rules. Elegant on paper. In practice everything it produced was on-brand and flat. "Meh" was the word my team kept reaching for. We had turned the brand into a box and asked for creativity inside it, and the box was the problem.

So we stopped generating from the rules. Now the system writes something interesting first, with room to be wrong, and only then runs the brand rules over the draft to catch what broke and fix those spots surgically, leaving the rest alone. Same rules, opposite end of the pipeline. Generated from, they produce compliance. Applied after, they produce work that's compliant and alive.

That inversion is the whole lesson. You can't encode taste, because the act of writing the rule is what moves taste out of its reach. Encode "strong hierarchy" and within a quarter everything has it and none of it is interesting, because the moment a quality becomes a checkable rule, clearing it stops being taste and becomes table stakes. An encoded standard holds last year's taste, frozen. So it belongs at the end, as a floor you clear once the work already exists.

Which means the goal most teams reach for, an AI that can tell good from bad, won't get them there. The leverage is elsewhere: take the judgment of the few people who have taste and make it reach a thousand decisions they'll never personally touch.

In practice that's a loop, and most of the noise about loops misses what makes one work. We wasted months trying to perfect the first draft. Most first drafts are 65-70% of the way there. The unlock was letting a person steer early and cheaply, explore angles, see a draft, change course, while a rubric defines where the loop is headed and each pass ends by clearing the floor.

And there's no single version of that loop that's right for everyone. Who's involved when, what you automate, what you measure, where the work pauses for a human call, what the interim artifacts look like, all of it has to flex to how your company actually works. The real thing you build is an orchestration layer you compose to fit how your company works, which is also why you can't buy taste infrastructure off a shelf. The judgment it scales is yours, and so is the workflow it lives inside.

The question isn't whether your system has taste. It's whether the bar it holds is higher this quarter than last.

New essay on how to build for that. Link in comments.