A catalogue is vocabulary, not authority
Pattern collections give you the shared words a room is already using — treating them as evidence about what works is the mistake that makes a designer sound well-read and wrong.
A designer from one of the portfolio companies is presenting an approval step in front of a billing action. Someone else says, “so this is a Governor.” Two people nod. The conversation moves on three times faster than it would have.
You want to be able to do that, and you want to be able to do it without acquiring a second, worse habit at the same time: treating the catalogue that supplied the word as though it had told you the pattern works.
This lesson uses one catalogue as the worked example, because it is the one most likely to be quoted at you. The skill is not about that catalogue. It is about the difference between a source that gives you names and a source that gives you findings, which is a distinction the room will collapse and you should not.
What Shape of AI actually is
Shape of AI is a free, curated library of AI interface patterns maintained by one named person, Emily Campbell, who also consults on AI design and product strategy. It is licensed CC BY-NC-SA. It sorts its patterns into seven types, and these are the site’s own definitions, verbatim:
| Pattern type | The site’s own definition |
|---|---|
| Wayfinders | “Help users construct their first prompt and get started.” |
| Prompt actions | “Different actions that users can direct AI to complete.” |
| Tuners | “Adjust contextual data, token weights, and input details to refine the prompt” |
| Governors | “Human-in-the-loop features to maintain user oversight and agency.” |
| Trust builders | “Give users confidence that the AI’s results are ethical, accurate, and trustworthy” |
| Identifiers | “Distinct qualities of AI that can be modified at the brand or model level to stand out.” |
| Dark matter | “Potentially nefarious, but certainly ambiguous patterns that impact user trust with questionable user value” |
Four things are missing from that site, and every one of them is a finding rather than a complaint. There is no stated methodology — nothing saying how a pattern gets in or what disqualifies one. There are no citations to research, anywhere. There is no date of any kind: no launch date, no last-updated stamp, no changelog, and no lastmod in the sitemap. The only year on the site is a 2025 copyright line. And there is one maintainer, rather than a team or an institution.
None of that makes the catalogue bad. It makes it a different kind of document from the CHI paper you read in the guidelines lesson, which printed its sample sizes, its four validation phases and its own scope limits. Shape of AI is one experienced practitioner writing down what she has seen shipped.
Do not quote a pattern count
The obvious thing to say in a room is “about sixty patterns across seven categories.” Say the seven, hedge the sixty, and here is why. This course’s own research pass fetched the same live homepage twice on the same day and got two different implied totals out of it, and both of those fetches silently omitted the Dark matter category altogether. The category is real: it was confirmed independently through the site’s own sitemap.xml and by fetching the category page directly, which returned the definition above word for word. If a number cannot survive being read twice, it is not a fact you own. It is a number you happened to see.
The one word worth taking from it today
Governors — “Human-in-the-loop features to maintain user oversight and agency” — is the closest published name for the thing you shipped. Your review gate over document extraction, with per-field confidence and the source crop sitting in front of every value before anything downstream fires, is a Governor in this vocabulary.
Using that word costs you nothing and buys you two things. The room stops spending thirty seconds working out what you built. And you have signalled that you have read the catalogue they read, which is the precondition for disagreeing with it later without sounding like you missed it.
What the word does not buy you is any claim about whether Governors work, when they are worth their friction, or what they cost. The site does not say, does not claim to say, and has no evidence on the page that would let it say.
The sentence that goes wrong, and the one that does not
These two sentences sound almost identical in a meeting, and only one of them survives being questioned:
- Goes wrong. “Shape of AI has a Governors category, so we should put a human in the loop before the payment fires.” The first clause is a fact about a website and the second is a design decision, and nothing connects them. Anyone who asks “based on what?” ends the argument, because the catalogue has no answer and you have borrowed its authority instead of bringing your own.
- Holds. “What we are describing is what Shape of AI calls a Governor — human-in-the-loop for oversight. That is a name, not evidence; the site states no methodology. The reason I think it is right here is what happened in our extraction review gate, which I can walk through.” The catalogue does the naming. You do the arguing. Nobody can take the second half away from you by discrediting the first.
That split — borrow the noun, supply your own verb — is the whole technique, and it generalises past this one site.
Four questions that place any catalogue in ninety seconds
This is this course’s own checklist rather than anybody’s published framework, and it exists because catalogues keep appearing and the room keeps treating each new one as authoritative for about a quarter. Ask:
- Who maintains it? One person, a team, an institution, or nobody any more. Shape of AI: one named person.
- What stands behind an entry? A study, a heuristic evaluation, or an observation that a shipped product does this. Shape of AI: observation of shipped products, with no methodology stated for how those observations were selected.
- When was it last touched? Shape of AI: unknown, and unknowable from the site. That is a real limitation for a catalogue of a field moving this fast, and it is the answer you give when someone asks whether a pattern is current.
- What does the publisher sell? Shape of AI is free and sells nothing directly, but its maintainer consults on AI design and product strategy, and a widely-cited catalogue is a credential. Naming that is not an accusation; it is the same reading you applied to Microsoft, Google and Apple in the first module.
Check your recall
Answer from memory — no scrolling back.
Retrieval check
You want to use the word “Governor” in a room without inheriting the catalogue’s authority. What has to be in the sentence?
Check your answer
A boundary. Something like: what Shape of AI calls a Governor — that is the shared name for it, and the catalogue is one practitioner’s observations with no stated methodology, so it tells us what to call this and nothing about whether it works. Here is why I think it works here.
The boundary is doing two jobs at once. It stops you from leaning on authority the source never claimed, and it pre-empts the person in the room who knows the site is one person’s catalogue and was about to say so. Getting there first is worth a surprising amount.
Hands on
Fill the Shape of AI row’s citation cell
Done when: The Shape of AI row in POSITIONS.md names one pattern category, quotes its definition verbatim from the site, and carries a one-line provenance note stating the maintainer, the absence of methodology and the absence of any date — each checkable by opening the site yourself.
- Open shapeof.ai and read all seven pattern-type pages. Do not work from this lesson’s table — the point of the exercise is that you have opened the source.
- In
learning/agentic-ux-canon/POSITIONS.md, fill the Shape of AI row’s first cell: the category and the verbatim definition of the one pattern you think is miscategorised, over-broad, or simply not a pattern. Governors is the obvious place to start, because it is the category your review gate lives in and you have opinions about what it lumps together. - Add the provenance line under it: maintainer, methodology, date. Three short clauses. If any of the three surprises you when you go looking, write down what you found rather than what you expected.
- Leave position, evidence and what would change my mind empty. The next two lessons are how those get filled, and filling them now from instinct is the habit this course exists to break.
- Bring the cell into the chat. I will check the definition against the site word for word, and I will ask you what the catalogue would have to publish before you would treat one of its entries as evidence.
What this does not cover
This lesson gives you a way to hold a source that supplies names and no findings. It does not give you the other half of the sentence — the part after “here is why I think it works here.” That is the next lesson, on grounding a disagreement in what you shipped, which takes your two real systems, states them at their actual size, and turns them into evidence a room will accept. After that, the third-why lesson stress-tests the resulting position, and the change-your-mind lesson asks what would falsify it.
Read this next — primary source
Shape of AIEmily Campbell — fetched 2026-09-05. One person’s curated catalogue of AI interface patterns, CC BY-NC-SA. Campbell also consults on AI design and product strategy, so the catalogue advertises the expertise she sells.
Spend twenty minutes clicking through the seven pattern types and read the entries under Governors and Dark matter closely. What you are collecting is words, not evidence: the site states no methodology, cites no research, and carries no dates. Read it the way you would read a dictionary — to find out what a room means when it says “governor”, not to find out whether governors work. Notice how quickly you start wanting to treat the entries as recommendations. That pull is the thing this lesson is about.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.