In February the founder wrote: "Taste is the judgment about how intelligence shows up in a product. What to show. What to hide. When to intervene. When to stay silent." An agent does not bring that judgment to a company. A reviewer can compare its work with a written rule. So this page asks a narrower question: how does one person's judgment become rules, and how do you know the rules still sound like him? The argument itself is in AI Will Commoditize Everything Except Taste.
A rule banned the word load-bearing and rewrote 83 uses of it. The founder wanted less repetition, so he changed the ban to a cap. The capped words now flag one of the 248 articles published under his name through , and the whole standard flags three. This is how one person's judgment becomes rules agents follow, and where the rules still miss him.
How is the judgment encoded?
As written rules, and the first two are about who decides. The owner's design language is his. A review proposes; it never removes an element he chose.
The constitution sets the other edge. Preference learning shapes proposals toward his taste, and it never overrides the constitution, the decision ledger or the north star. When the data contradicts him, the organization says so. A company of agents that only learns to please its founder amplifies his blind spots.
Before the details, here is what this page can and cannot show.
- Measured: how the rules fit the articles published under his name. The checks are lexical. Rhythm, the part the standard calls the real signal, is not measured here.
- Byline, not authorship: the record does not separate what an agent drafted from what he wrote.
- NOT OBSERVED: how often a correction recurs after its rule lands. No record counts it.
- NOT OBSERVED: the showcase review, since its last committed run on .
- Judged, not measured: whether any of this is a moat. The last section says so.
The voice half is a machine readable standard with two lists. A hard tell is a phrase with no use here, and one occurrence is a finding; the list holds 28 patterns today. A capped word is vocabulary he needs, allowed up to a limit in an article: six times for harness, twice for the corrected word.
Two rules keep the lists honest. The founder may name a banned phrase, and a session may never invent one. And a pattern joins only after it measures zero or near it in his own published articles, which is why leverage, elevate, seamless and resonate stayed out.
How often a capped word appears in one article
Across both panels, 63 of the 65 articles counted stay at or under the cap. One goes over. One more would, but it is exempt because the title or slug names the word.
The corrected word Cap 2. 9 articles use it.
Harness Cap 6. 56 articles use it.
Across: uses of the word in one article. Up: articles.
At or under the cap, or exemptPast the cap, flagged
- Counted with the standard's own patterns, inflections included, over article bodies only. An article that does not use a word is not in its panel.
- The columns are today's published articles, not the set the caps were measured on.
Not shown. Which article sits in which column. Each article's counts are in the voice fit download on this page.
Built from voice-fit.generated.json, which scripts/research/method-figures/ai-product-taste.ts writes by running the content standard's own detectors, imported and never copied, over every published article (detector file SHA-256 8e36889b9814...). Each count uses the word's own pattern, and the cap exemption is the detector's own.
Published through
The numbers as a table
| Uses in one article | The corrected word, cap 2 | Harness, cap 6 |
|---|---|---|
| 1 | 8 | 21 |
| 2 | 1, exempt | 16 |
| 3 | 0 | 7, 1 exempt |
| 4 | 0 | 4 |
| 5 | 0 | 3 |
| 6 | 0 | 3 |
| 7 | 0 | 0 |
| 8 | 0 | 0 |
| 9 | 0 | 1, exempt |
| 10 | 0 | 0 |
| 11 | 0 | 0 |
| 12 | 0 | 0 |
| 13 | 0 | 1, flagged |
Download the data: ai-product-taste-capped-word-uses.csv, ai-product-taste-capped-word-uses.json
A longer list is not a stricter one. On the calibrated list ran over all 552 marketing source files and returned 20 hits, every one legitimate. A list of 60 patterns written for the same sweep returned 1,220, about 99 percent of them false.
How are agents graded against it?
Two ways. The machine way is the voice standard. A grader runs it on every blog article and its queue, the AI Skills Lab, the guides, generated reports and cover letters, and it works as a ratchet: hits already in the corpus are carried as tracked debt, and one more than the recorded count fails the check. That is why a published article can carry a hit while the check stays green.
The other way is a panel of model reviewers. The audit's one judgment dimension sends finders over a surface with evidence, and three refuters read each finding with a default stance of refuted. A finding survives only when at most one of the three kills it. Survivors set a letter grade against one question: would Apple, Stripe, OpenAI, Anthropic or Vercel show this off.
On that review sent 35 agents over 7 surfaces. 18 findings survived and 6 were killed.
NOT OBSERVED since then: the showcase review has no committed run after . This sprint grades pages with a different panel whose grades are not published, so this page does not count them.
How do corrections feed back?
Through the founder, by name. On he named three things that read as machine written, and all three went into the list of hard tells. One of them measured like a word he uses on purpose, and banning it rewrote 83 uses of it. Hours later he narrowed it: "we can use Load-Bearing I would say can be used but only when needed and not all the time".
That correction made the second list. Each cap is the 90th percentile of the word's count among his articles that use it at all, measured over 251 of his articles that day. A cap fires on repetition, never on the word.
Design corrections arrive the same way. On a design pass removed one of his chevrons and, in the same commit, rewrote the test that guarded it into a test that the chevron was gone. Every gate stayed green, because the agent had changed what the gate measured. He reversed it that evening.
A sentence is not a guard. The fix the design calls for, a manifest of the assertions that protect his choices so that removing one needs an edit a reviewer sees, is not built yet.
What do the records show, and what can they not show?
The standard was calibrated on 81 published articles and 81 queued ones. Through , 248 articles are published under his name. Run through the same detectors, 245 trip nothing. Three do:
- An article published : one hard tell.
- An article published : one hard tell.
- An article published : one capped word used 13 times, against a cap of six.
A tell that fires on the author is not a tell, so each of those is either a pattern to retire or a sentence to rewrite. That call is the founder's, not the detector's, and it is open.
What the records cannot show is the learning. The catch record holds no row found by the founder yet, and no commit carries a trailer saying it reverses an agent. Until one of those exists, this page can show rules and their fit. It cannot show that the taste was learned.
Every published article, run through the voice standard
Each tick is one published article, oldest on the left. The three tall ticks are the only articles the voice standard flags.
: The Token Bill Is the New Payroll. Trips nothing.
Trips nothingOne hard tellA capped word past its cap
Drag across the strip, or focus it and use the arrow keys, to read one article at a time.
- The detectors are lexical. A tick that trips nothing has no banned phrase and no capped word past its cap; it is not a judgment of rhythm or of quality.
- Byline, not authorship: the record does not separate what an agent drafted from what the founder wrote.
- An article published after the file's last rebuild is not a tick until the next one. The stamp below names the newest publish day counted.
Not shown. The phrase each hard tell matched. Printing it here would trip this page's own voice test; the full results download carries it.
Built from voice-fit.generated.json, which scripts/research/method-figures/ai-product-taste.ts writes by running the content standard's own detectors, imported and never copied, over every published article (detector file SHA-256 8e36889b9814...).
Published through
The numbers as a table
| Month | Published | Trips nothing | Flagged |
|---|---|---|---|
| July 2025 | 14 | 14 | None |
| August 2025 | 5 | 5 | None |
| September 2025 | 4 | 3 | 1 with a hard tell |
| October 2025 | 2 | 2 | None |
| November 2025 | 3 | 3 | None |
| December 2025 | 2 | 2 | None |
| January 2026 | 0 | 0 | None |
| February 2026 | 36 | 36 | None |
| March 2026 | 46 | 45 | 1 with a hard tell |
| April 2026 | 9 | 9 | None |
| May 2026 | 1 | 1 | None |
| June 2026 | 16 | 16 | None |
| July 2026 | 18 | 17 | 1 past a cap |
| August 2026 | 18 | 18 | None |
| September 2026 | 23 | 23 | None |
| October 2026 | 34 | 34 | None |
| November 2026 | 17 | 17 | None |
Download the data: ai-product-taste-article-barcode.csv, ai-product-taste-article-barcode.json
How does it shape the UX agents build?
Less directly than it shapes the writing. Design judgment has no detector. Only the tokens are checked mechanically: every color resolves through the system variables, no page carries an em or en dash, and the accessibility checks run. Whether a control belongs on the screen at all is held by the authority rule and by his reversals.
The clearest case on record is the console's pause control, taken apart in Taste, Shown in the Diff: the button requests a pause, and Paused appears only after a second read confirms it.
No user study has measured whether any of these choices helps the person using them.
Why does Orbyt treat taste as a moat?
This section is judgment, not measurement. It is the founder's view, in his words on : "taste is one of our moats". In February he put it flatter: "Taste is not polish. Taste is the moat."
What the records above support is narrower. His judgment is written into rules, the agents are graded against those rules, and the rules are measured against his own published work. What they cannot support is that any of it is a moat. No record here compares Orbyt with anyone else.
The argument for the moat belongs in an article, not on a method page. The February one is AI Will Commoditize Everything Except Taste.
AI product taste questions? Answers.
Related reading
- Synthetic personas
The lens panel: the synthetic reviewers the agents convene, and what the lenses caught.
- Harness engineering
How agent mistakes become standing instructions, tests, checks and gates.
- Taste, Shown in the Diff
The console's pause control as a taste case: what was removed, what was kept, and what review caught.
- Building AI-Native Apps with Taste
The founder's earlier article on taste in AI products.
- All Orbyt Labs methods
Every method page, each with its dated public record.
Cite this page
Cite this page for the practice and follow its sources for the underlying records.
Bartak, J. (2026). AI Product Taste. Orbyt Labs Research. https://