The field – Real systems, and who owns the code – The System – Decisions made once, used everywhere

The systems worth knowing aren’t five versions of the same thing – each one is a different answer to a different problem, and reading which problem is the transferable skill. Dates matter throughout: this landscape moved more in 2025–26 than in the five years before it.

Five philosophies

Material 3 Expressive (Google) – the consumer-scale, research-led one. Announced May 2025, stable 3 September 2025, and unusually for a design language, it shipped with receipts: 46 internal studies and 18,000+ participants, with Google’s research page claiming users spotted key UI elements “up to four times faster” in the expressive layouts. The most rigorously token-driven of the majors, open source, cross-platform. When someone says “systematised at scale”, this is the reference point.

Apple’s Human Interface Guidelines, rewritten around Liquid Glass – the material-and-motion one. Announced 9 June 2025 and widely described as the biggest shift since iOS 7 (commentators’ framing, not Apple’s), it landed simultaneously across every Apple platform for the first time. Glass as a hierarchy tool – foreground controls elevated above content – not decoration. Then the part the industry got to watch live: a year of user pushback on transparency, and at WWDC 2026 (8 June) Apple announced a settings slider running “from ultra clear to fully tinted”, plus readability changes, folded into iOS 27. A confident flagship bet, user-tested in the wild, corrected in public – even at Apple’s scale, systems are hypotheses.

Shopify Polaris – the distribution experiment. On 1 October 2025 Shopify shipped Polaris Web Components, framework-agnostic and loaded straight from Shopify’s CDN – apps inherit design-system updates automatically, no code change – while the old React library is now deprecated and in maintenance mode. A deliberate bet that “works with any framework or none” outlasts any framework’s popularity cycle.

IBM Carbon – the enterprise specialist. Data-dense components, strong accessibility defaults, governance built for a large organisation’s many teams; v11 current (v10 fully deprecated 30 September 2024), v12 in Preview behind feature flags as of mid-2026. Carbon exists to prove the point that a system is built for a problem: put Material’s expressive shapes in a compliance dashboard and you’ve solved the wrong one.

Atlassian – the one that says out loud where this is heading: as of mid-2026 its own blog describes the system as becoming “a context engine for the AI era”, with motion tokens structured so “both humans and AI can understand the intent behind each animation”. Marketing language about a strategy, so hold it as stated intent – but the direction matches everything else on this page: systems increasingly have a third audience, and it reads tokens, not Figma frames.

The one that isn’t a system – and why it matters most

shadcn/ui distributes components by copying source code into your project – code ownership transfers at install time, and there’s no library boundary left to negotiate with. Feel the difference a folder makes:

Two identical buttons – try to change each one

Same component, different homes. The home decides who may edit it.

Installed package

🔒node_modules/ui-kit/Button.js

export function Button() {
  return <button style={{
    background: indigo,
    borderRadius: rounded,
  }}>Get started</button>
}

Get started

Copied into your repo

src/components/Button.tsx

export function Button() {
  return <button style={{
    background: indigo,
    borderRadius: rounded,
  }}>Get started</button>
}

Get started

Both buttons render identically today. The difference only appears the moment you want something changed.

The locked half is the traditional model: components arrive as a dependency, every customisation someone might want has to be pre-built into the library's settings, and anything else means waiting on the maintainers. The editable half is the shadcn model – the source was copied into your repo at install time, so there's no boundary left to negotiate, and (not incidentally) an AI agent can read and modify it like any other file in the project. One caution to carry out of this page: what you get is a distribution model and a starting point, deliberately styled thin – the coherence and governance that make a design system are still yours to build.

Scale check, with dates: the January 2025 Rising Stars report named shadcn/ui the most popular JavaScript project of 2024 by new GitHub stars – ahead of every framework and runtime. The following year’s report has it third overall (n8n took first by a wide margin) while still first in the component-library category – so “dominant in its lane, no longer the headline” is the current truth, and the AI-tools connection is the durable part: v0, Bolt, Lovable and Cursor default to shadcn-shaped code because plain, visible source is what agents read and modify reliably. Registries extend the model – any team can publish a component registry and pull from it via CLI. The precision worth carrying: shadcn is a distribution model and a starting point, deliberately styled thin so you customise it into your system. Adopting it doesn’t hand you a design system – the coherence and governance are still yours to build.

Reading a system’s pulse

Last habit, and the course’s closing honesty check: a system’s public open-source presence and its internal organisational health are different things, and they can diverge completely. The standing example:

Case studyUber BaseThe design system serving thousands of Uber engineers – whose public open-source mirror deliberately went quiet. Two health signals, easy to conflate.This course ends on how systems live and age – Base is the caution against reading a repo's community activity as the system's pulse, in either direction.

Two names, deliberately distinct: Base is Uber’s company-wide design system – per Uber’s own documentation it “defines the foundations of user interfaces across Uber’s ecosystem of products & services”. Base Web is the open-source React implementation of it, published to npm as baseui. They are not the same thing, and as of mid-2026 they have genuinely different public temperatures – which is the teaching point.

Surface signals first: the GitHub repository is not archived, commits land regularly (as recently as 7 July 2026 at the last check), and npm versions keep shipping (18.2.0 published 2 July 2026). Yet the README carries a standing notice, in place since December 2023: “We are limiting our engagement with this repository while still mirroring our internal development.” Read precisely, that means Uber’s engineers keep developing Base Web – internal work is mirrored out – but community-facing maintenance (issues, external pull requests) has been explicitly deprioritised for years.

Meanwhile the internal story points the other way entirely. Uber’s March 2026 engineering blog describes Base as the system that “serves thousands of engineers” and details fresh investment: uSpec, an agent-driven tool that connects to Figma and generates component specification pages in minutes rather than weeks, with a live roadmap behind it. The honest verdict: not abandoned, not thriving-in-public – a system whose internal organisational health and public open-source presence have diverged. Both of the easy conclusions from a surface check (“quiet repo, dead system” and “shipping releases, all is well”) would be wrong here, in opposite directions. Check both signals before betting on either.

The moves, named – follow one into the glossary:

The system vs its open-source implementation
Reduced external engagement is not abandonment
Specs generated by agents – minutes, not weeks
Check both signals before declaring health

Recap

Systems are answers to problems – Material to consumer scale, Carbon to enterprise density, Polaris to distribution, Apple to material identity. Name the problem before judging the answer.
Even flagship bets get corrected in public – Apple announced Liquid Glass's transparency slider one year in. Systems are hypotheses with release dates.
shadcn/ui is a distribution model, not a design system – code ownership transfers to you, and so does the system-building work.
Check both signals before declaring a system alive or dead – Uber Base's public repo went quiet in exactly the years its internal investment grew.

Terms you now owncode ownershipcomponent registry

In a review “Which problem was this system built for – and is it ours? A beautiful answer to someone else's problem is still the wrong answer.”

To an agent “Use shadcn-style copied-in components as the starting point, then apply our tokens and conventions on top – and tell me which parts are still the generic defaults so I know what's left to make ours.”

This page practises it:A field guide is only as good as its dates – every claim on this page carries one, because the survey's own lesson is that this landscape moves and currency is the point.