Uber says Base defines the foundations of user interfaces across its products and services, and brings all Uber experiences together under a single, unified framework.
The section overviews are public. Detail pages, including Design tokens, Color 2.0, What's new and About Base 2.0, ask for Uber's staff login, so Shortcut describes the structure and not the guidance inside.
Atlassian describes it as a collection of design guidelines, foundations, tools and components, built and supported by its developers, designers and content designers.
Google's open-source design system. Material 3 is the current version, and Material 3 Expressive is its 2025 update: new components, spring-based motion and stronger emphasis, without replacing Material 3.
Shortcut has read the Overview page of all 36 components, the Breakpoints page and the Expressive announcement. It has not read Material's pages on tokens, colour, typography, shape or elevation, nor any component's Specs or Accessibility tab.
Named for completeness. Shortcut says nothing about them until it has read their own pages.
Shopify
Polaris
Public design system
Shopify's UI framework for apps on its platform.
The old Polaris site now redirects to a developer reference for web components. Shortcut found no design guidance pages there, so Polaris is not compared.
Companies whose published design material is worth studying. None publishes a documented design system Shortcut could verify, so none is described as having one. Gojek names its components but gives no guidance for them.
Grab
Grab product design
Product design reference
Grab publishes articles on how its designers and researchers work. Shortcut found no public Grab design system, so this is a reference to its published practice, not to a system.
Three of the four articles used here date from 2019 to 2021. The product has changed since; the reasoning is what is worth reading.
Granola publishes dated product updates. Shortcut found no public Granola design system, so this is a reference to its release notes, not to a system.
Only refinements that Granola's own updates page describes are listed. It does not go into detail on keyboard behaviour, text wrapping or icons, so none of that is claimed here.
Gojek's design site shows its brand: illustration, motion, tone of voice, typography and a list of design system components. It names the components but publishes no usage guidance for them, so Shortcut treats it as a brand reference, not a design system.
The site is a showcase more than documentation. Several pages are mostly images, which Shortcut cannot read as text, and it does not describe what the images show. The footer is dated 2025; no page gives a publication date.
What each system itself says about viewports, touch and devices, on the pages Shortcut has read. No values are filled in for a system that does not publish them. Compare by viewport and platform
Base
Responsive guidance behind staff login
The pages that would cover this need a staff login, so Shortcut cannot say what the system publishes.
Atlassian Design System
Breakpoints, grid, content width
What it says (6 points)
Atlassian
xxs 320 to 479, xs 480 to 767, s 768 to 1023, m 1024 to 1439, l 1440 to 1767, xl 1768 px and above.
Fixed-wide grids have a maximum width of 1296 px and are the default. Fixed-narrow grids, at 864 px, are for long-form reading. Fluid grids have no maximum and should be used sparingly.
Unfolded in landscape is like a tablet: a two-pane layout with a navigation rail. Folded is like a phone: a single column with a bottom navigation bar.
Uber: an agent that writes component specs in Figma
The design-system lesson
Craft guidance
A design system is no longer only documentation for people. Uber's agent is useful because the system underneath is structured: real token names, real variants, a spec template. The AI contributes interpretation; the structure contributes accuracy. Without the second, the first has nothing to stand on.
What they say they did (6 points)
Uber's design systems team built uSpec, which connects an AI agent in Cursor to Figma through the open-source Figma Console MCP.
The agent reads the real component: structure, tokens, variables and styles. It then writes finished spec pages into the Figma file.
Each part of a spec has its own agent skill: anatomy, API, properties, colour annotation, structure and screen reader.
The screen-reader skill loads platform accessibility references first, so the agent selects from documented properties and does not guess.
Uber says a screen-reader spec covering iOS, Android and web generates in under two minutes, and that the pipeline runs locally so design data stays inside its network.
Atlassian: the design system as structured content for AI
The design-system lesson
Craft guidance
Design systems used to have two readers: designers and developers. AI agents are a third, and they cannot skim a Figma file or watch a video. Guidance that is not written down in a structured form gets replaced with something generic. Atlassian's line: documentation that drifts from the code is worse than none.
What they say they did (6 points)
Atlassian found its design-system guidance spread across documentation sites, Confluence, Figma files, videos and code, some of it outdated and most of it written for people to look at.
It broke the guidance into consistent, machine-readable chunks with schemas for components, icons, tokens, lint rules and foundations, kept alongside the code.
Each schema covers usage, code examples, props, content standards and accessibility requirements.
The same source generates its MCP server, its agent skill and its DESIGN.md files.
Atlassian reports that agents using it produced 4.9% more accurate code with 11% fewer errors and finished tasks 34% faster than agents without it.
It says unstructured sources led to outdated patterns, missed accessibility requirements and invented components.
Generated UI looks generic when the model knows nothing about your brand, components or spacing. A portable design file fixes the look cheaply, which helps prototypes. It does not make a model use your real components, so alone it can produce a convincing second system. That is an AI-look signal by a more polished route.
What they say they did (7 points)
DESIGN.md is an open Markdown format, designed by Google for its Stitch tool, that Atlassian describes as a portable snapshot of a team's brand and UI patterns.
The first part lists design tokens in machine-readable form. The second is guidance for people and agents on colour, spacing, layout, elevation and components.
Atlassian's summary of the problem it addresses: "Generic in, generic out."
In a one-shot demo in Figma Make, the file moved output from generic to recognisably Atlassian.
In a production test it used about 92% more tokens than the baseline and varied more between runs. Atlassian says these results are not conclusive.
Agents given DESIGN.md were more likely to re-create components than use the existing library.
Atlassian suggests it for art direction, quick prototyping and theming, and MCP servers or skills with lint rules for production work.
Atlassian: three ways to hand a design system to an agent
The design-system lesson
Craft guidance
How the design system reaches the model is now a design-system decision, with costs and trade-offs. Designers need not build it. They should ask which route their team's tools use, because it decides whether generated UI follows the system.
What they say they did (6 points)
Atlassian delivers its design system to AI agents three ways: an agent skill, an MCP server and a command-line tool.
All three draw on one structured source, so they stay in step.
The context covers component APIs and usage guidance, tokens, icons, accessibility guidance and lint rules.
It added the command-line tool because many agents have no MCP client but can run a command.
Atlassian reports tasks finished about 8% faster with about 8% fewer tokens using it.
Its advice includes keeping these interfaces in sync, and reading agent transcripts alongside the metrics.
AI folded into an existing design practice, not set up beside it. Two things to copy: the tool was trained on Grab's own visual language, and the designer's job moved toward judging output. The figures are Grab's own, from June 2025.
What they say they did (6 points)
Grab's Head of Design describes a deliberate, hybrid approach: a vision set from the top and experimentation from the bottom, with teams given specific design challenges to solve with AI.
It built Mosaic, an internal image generator trained on Grab's visual identity. Grab says an illustration that took a day now takes about 15 seconds, and that it produces around 800 a day.
Illustrators now spend their time writing detailed prompts and refining outputs.
Designers use agentic coding tools to build prototypes with working logic, navigation, form validation and mock data.
Other tools generate usability-test scripts and flag usability issues.
The article stresses human oversight, and notes that models trained on past data are weaker at predicting future trends or reactions to new technology.