Reference · 2026

The glossary

Every term from The Adaptive Brand and its three tutorials, in plain language, for the designer or PM who just got linked into the middle of it. No prior reading required. Each entry says where the idea lives, so you can go deeper when one earns your attention.

Kyle Cooney, Head of Product Design · The full article · The tutorials

One sentence before the definitions, because every term below hangs off it: your product's screens and sentences are increasingly generated by AI, and the adaptive design system is the set of files and checks that lets them adapt to every person and channel while the things you must never get wrong stay never gotten wrong. That is the whole idea. The vocabulary just gives its parts names.

The big idea

Adaptive design system: A design system that governs AI as well as people. A classic design system tells designers which buttons and colors to use. An adaptive one also feeds the AI that builds your screens and the AI that talks to your customers, and checks their work against your rules. The short definition: it lets generators adapt every surface while it holds the commitments.In the corpus: The article, Preface
Adaptive brand: A brand that stays recognizably itself while AI changes everything else: the words, the layouts, the language, the channel. The quality the system exists to produce.In the corpus: The article
Coherence (vs. consistency): Consistency is every button looking the same. Coherence is your product feeling like one product even when no two customers see the same screen. Old design systems needed consistency; generated products need coherence.In the corpus: Ch 14, The Discipline
The two generators: The two AIs now making your product. At design time, an agent in your codebase builds the screens. At runtime, a model inside the product writes the sentences and, increasingly, composes the interface. Both are fluent. Neither knows what your product must never say, which is why the system exists.In the corpus: Ch 1, The Bottleneck Was the Brand
Design time / runtime: Design time is while your team builds the product. Runtime is while a customer uses it. The same rules have to hold at both moments, which is why the system gets built twice in the same shape.In the corpus: Ch 8, The Control Plane, Twice
Agent / harness: The agent is the AI worker that reads your files, writes code, and runs commands when you ask in plain English. The harness is the app it works inside: Claude Code, Codex, Cursor. Everything of value lives in your files, not the harness, so you can switch harnesses without losing anything.In the corpus: Getting started, step 1

The files

The system is, in the end, a small set of plain-text files in your repository. These are the ones with names.

Brand file (brand.md): One plain-text file that describes your brand specifically enough for a machine to follow it. “Warm but professional” is a vibe; the brand file is rules: exact colors, type, tone by situation, and the words you never use.In the corpus: Getting started, step 3
Commitments file: The short list of things that must stay true no matter who or what wrote the sentence: what you always disclose, what you never claim, which actions need the customer's explicit yes. Every other check in the system traces back to this file.In the corpus: Part I, step 2
Never-say list: The sentences your product will not produce. Ever. In any language, on any surface, no matter how natural they would sound. Each entry is a commitment, not a style preference.In the corpus: Ch 6, Brand as Governed Context
Instruction file (AGENTS.md): The standing brief the agent reads at the start of every session: what the project is, where things live, and the rules it must never break. An open format that every major AI coding tool reads.In the corpus: Getting started, step 4
Skill (skill file): One plain-text file per domain of judgment: brand, copy voice, accessibility, components. When you catch yourself explaining the same thing to the agent twice, that is a skill file asking to be written.In the corpus: Getting started, step 4
Decision log: Two or three sentences recorded at the moment of every real decision: what you chose, what you rejected, and why. The agent reads it before generating, so settled questions stay settled. Six months of entries is better onboarding than any deck.In the corpus: Ch 11, The Memory of the System
Design tokens: Your visual decisions stored as named values: colors, spacing, type sizes. Change a token and every screen that uses it follows. The single source of truth for how the product looks.In the corpus: Ch 4, One Source of Truth
Semantic tokens: The token layer that names meaning instead of appearance: “warning” rather than “orange-600.” Generators touch only this layer, which is how a thousand generated screens stay siblings and a rebrand stays an afternoon.In the corpus: Ch 4

The unit of design

Behavior: What a moment of the product is for, named as a verb: Explain, Confirm, Compare, Warn, Repair. A behavior is designed once, without reference to any screen, then rendered as a chat message, a form, an email, or a voice turn. Screens change per person and channel; the behavior is what stays designed.In the corpus: Ch 5, Semantic Components
Semantic components: The article's name for behaviors as reusable objects. Classic component libraries standardized the nouns (Button, Card, Modal); semantic components standardize the verbs one level up.In the corpus: Ch 5
Product grammar: The stable underlying patterns of your product category: transfers, listings, carts, statements. Products with a settled grammar can be modeled deeply; products still finding their shape should be prototyped freely and modeled lightly.In the corpus: Ch 3, The Adoption Threshold
Profile (conversation profile): How the same behavior adapts per market or person: pacing, formality, which examples land, how consent is asked. The words and rhythm move with the profile. The disclosures move with nothing.In the corpus: Ch 7, Every Surface, One Promise

The checks

The half no style guide ever had: machinery that verifies the generators' work instead of hoping someone reviews it.

Contract: Machine-checkable bounds on what a behavior may say: which facts it may state, which disclosures it must include, which claims it may never make, and what happens when a generation breaks the rules. The brand file teaches the generator; the contract checks its work.In the corpus: Ch 6
Renderer contract: Placement guarantees for every surface, in terms of what a person actually encounters: the fee is visible before the confirm button can be reached, cancel is one step away, the warning is as prominent as the risk. Like accessibility testing: not “is the disclosure somewhere,” but “will the customer meet it.”In the corpus: Ch 7
Fallback (degrade-to-template): The plain, pre-written version of a screen or message that ships when generation fails or fails its checks. It will be boring. It will also be correct, and boring-but-correct is the right floor for anything that moves money or collects consent.In the corpus: Part I, step 7
Eval: Taste written down as a test: a rule precise enough to check across a hundred generated variants. “Fail if the fee is not stated before the consent control, in any phrasing.” Metrics decide winners; evals decide what is allowed to compete at all.In the corpus: Ch 10, Evaluation-Driven Design
Golden set: The ten to twenty example scenarios you always test against: the typical customer, the tricky one, the delayed transfer. Saved as files, so every test run measures against the same bar.In the corpus: Part II, step 9
Eval harness: One command that generates fresh output for every golden scenario, runs every check, and keeps score over time. What makes “retest everything when a new model ships” a command instead of a week of work.In the corpus: Part II, step 9
Hook (vs. instruction): An instruction can be forgotten; a hook fires every time. A hook is an automatic check wired into the workflow, like blocking any change that hardcodes a color. Preferences belong in instructions. Rules you cannot afford to forget belong in hooks.In the corpus: Ch 8
Floors and dials: Floors never move: the disclosure that always appears, the claim that is never made. Dials are what you tune: tone, conversion, delight. The fastest way to raise a dial is usually to quietly lower a floor, which is why the system never lets the two argue.In the corpus: Ch 10
Authority: Which actions the AI may take on its own and which need a human or the customer's explicit consent. Sending the email is not the same kind of act as drafting it.In the corpus: Ch 8
Traces: The replay: a record of what the model knew, said, and did on each turn, so when someone asks “why did it say that,” the answer takes a minute instead of a meeting.In the corpus: Ch 8

The practice

Control plane: The layer around the AI that decides what it gets told, what it may do, how its output is checked, and what evidence is kept. Built twice in the same shape: once around the agent in your codebase, once around the model in your product. The AI is the tenant; the plane is the building.In the corpus: Ch 8
Governed context: What the AI is told before it works, curated instead of scraped: the brand file, the commitments, the behavior, each admitted because it changes a decision and each traceable to a source (its provenance). The system also declares what it does not know, so the model falls back instead of improvising.In the corpus: Ch 6 and Ch 8
Exposure (and the ladder): How much governance you need, as two numbers multiplied: the cost of one wrong sentence, times the number of surfaces that can produce one. Adopt in rungs (brand file, then tokens, then behaviors, then contracts, then evals, then traces) and stop where your exposure stops.In the corpus: Ch 3
The three renderings: The canvas (for looking), the prototype (for feeling), and the production app (for real), all generated from the same source so they cannot quietly drift apart. The design file, the demo, and the product stop being three things to keep in sync.In the corpus: Ch 4
Experiment lane: How you break your own rules on purpose. An experiment names which dials it suspends, where it runs, and when it ends. It may suspend any dial. It may never suspend a floor. If it wins, it becomes a named pattern with checks of its own.In the corpus: Ch 10
Truth loop: The rule that keeps the system current without relying on diligence: a change to the product that does not update the system does not ship. New behavior, new check. New market rule, new profile entry. Same change, same moment.In the corpus: Ch 11
Scaffolding vs. commitments: Scaffolding compensates for what today's models cannot yet do reliably, and it should be retired as they improve. Commitments are choices no model improvement makes for you: what you promise, disclose, and never claim. Confusing the two is how teams either over-build or under-protect.In the corpus: Ch 13, The Case Against
Sunset review: The scheduled check, every quarter and after every major model release, asking of each piece of scaffolding: is the limitation this compensates for still real? Retire what is not. Keep the commitments; they were never scaffolding.In the corpus: Part I, step 12
The Bitter Lesson: A famous machine-learning essay warning that handcrafted workarounds eventually lose to better models and more compute. The reason the sunset review exists: build what survives model progress, and plan the retirement of what merely compensates for it.In the corpus: Ch 13

Generative UI

The newest surface: interfaces the model composes on the fly, and the vocabulary for keeping them governed.

Generative UI: Screens the model composes in the moment, per person and per situation, instead of screens a designer drew in advance. The screen becomes an output; what you design is the vocabulary it composes from and the rules it cannot break.In the corpus: Part II
Catalog: The fixed list of components the model is allowed to build screens from, with exactly the inputs each accepts. The model cannot invent a part that is not in the catalog, which is the first line of defense.In the corpus: Part II, step 3
Registry: The real code components the catalog's names map to, styled by your tokens. The catalog is the menu; the registry is the kitchen.In the corpus: Part II, step 1
Spec: The generated screen as data: a small JSON description of which components appear and in what order, which a renderer then draws. Because a spec is data, every promise about the screen can be checked in code before a customer sees it.In the corpus: Part II, step 1
State path: How facts reach a generated screen: by reference to your real data (“show the fee at transfer.fee”), never as text the model wrote. The model composes the screen but cannot write the numbers, which retires the whole category of invented-fact failures.In the corpus: Part II, step 3

Missing a term, or found one of these unclear? The definitions live in one file; tell me and it changes. For the argument behind the vocabulary, read the article. To build the thing itself, start with Getting started.