Ask Claude to build a screen and it will build one that works. The buttons do what they say, nothing crashes, and every check passes. Then you use it and something is off. The words sound like a manual. The flow asks one question too many. It works, and you'd never choose to use it.
Checks that test whether something works can't see this, because nothing is broken. So we review feel separately, with reviewers whose only job is to ask how it feels.
A separate review for feel
Every piece of work a player will see goes through a taste review before it's marked done. It covers the words, the visuals, how the flow moves, and whether the app feels like itself. The reviewers are AI models from three different companies, because each one has its own blind spots and its own taste, and we want those to cancel out rather than stack up.
If any reviewer says "rethink", the work stops and all three opinions come to me. Working code doesn't override it.
The reviewers judge what a player gets: the screenshots, the words on them, the order things happen in, and the stranger test, whether someone with no context gets it in three seconds. Claude, asked about its own work, tends to explain why it's fine. A reviewer that only judges the result has nothing to defend.
It read like a rulebook
Here's the kind of thing it catches. In our party game, every secret role comes with a short explanation for the player who gets it. Claude's first draft of one was accurate, complete and clear. Both outside reviewers sent it back with the same diagnosis: rulebook prose.
That's Claude's natural voice when nobody steers it. It writes like documentation, complete and careful and neutral, because most of what it has read that explains how things work is documentation. For a manual, that's ideal. For a party game it's a small disaster that every check will wave through.
They were right. It read like the rules insert in a board game box, in a game that's supposed to sound like the friends around the table. Once it was rewritten in that voice, the same information passed easily. No test could have flagged it. Nothing in it was wrong.
The reviewers are wrong too
Taste reviewers are tools, and tools misfire. When we asked one to judge a set of 15 small emblems, it judged them as if they were a single app icon and came back with "do not ship this as the app icon". It was answering a question nobody asked.
Run through a reviewer built to judge a set as a family, the same emblems passed, and that reviewer independently flagged the one real flaw we'd already noticed. Before trusting a harsh verdict, check that the reviewer was asked the right question.
Some reviewers are simply harsh. One of our three, Grok, kept marking down the personality and delight of one app's opening screens, even when we handed it wording taken from its own earlier suggestion. The rule for that case: take its reasoning seriously for a round or two, then write down that it disagrees and move on. Gutting good copy to please one reviewer makes the app worse.
Where AI taste runs out
The clearest example came from somewhere other than the apps. When we started writing posts about the studio, AI judges scored every draft. They gave scores of 7 to 10 to posts I killed on sight, as boring, as a diary, as a small problem blown up into an essay.
The judges weren't careless. They were scoring against a sensible list: is there a story, is there a point, is it clear. The killed posts ticked every box and were still dull, because "would anyone stop scrolling for this?" isn't on any list.
In the second round I killed 11 of 12 drafts the judges were happy with. So for anything where the point is whether a person will care, there's no AI judge at all anymore. I strike what doesn't work, and that's the only filter. The page you're reading went through the same process.
AI taste reviews are good at catching things that are clearly off, like rulebook prose or a cluttered screen. They're bad at knowing whether something is interesting. That call stays with a person.
If you build with Claude, try this
- Use it before you call it done. Tap through the real thing the way a stranger would.
- Ask a separate question about feel. "Does this work?" and "Does this feel right?" need different reviews. Ask Claude for the second one in a fresh chat, so it isn't defending its own work.
- Get more than one opinion. Different AI models have different taste. Agreement means more than any single score.
- Check the reviewer's question. A harsh verdict on the wrong question is noise.
- Keep the last word. If you wouldn't use it, it isn't done, whatever the scores say.
Working is the minimum. Whether someone wants to use it is the actual job.