expo-design-system

Expo Design Systems

Make every screen in an app draw from one visual source of truth: a token theme and a small set of reusable components. This skill defines where tokens live, what they cover, how reusable components are shaped, and when a repeated view earns promotion into the system.

Sibling skills own the layers around this one:

• expo-native-ui - platform styling rules (HIG, semantic colors, controls, shadows syntax). Follow it for what values look native; follow this skill for where values live and how they're reused.
• expo-project-structure - folder skeleton for new apps.

For Tailwind projects, keep tokens in global.css as CSS variables and follow the styling library's own setup guidance. The scales and naming in this skill still apply; only the storage format changes.

References

Consult these resources as needed:

references/
  audit.md        Audit an existing app for design-system drift: grep checks,
                  scoring rubric, incremental adoption plan, and templates for
                  documenting or extending components
  native-slop.md  The 20 named anti-pattern tells of AI-generated apps (The Web
                  Modal, Everything's a Card, ...) with grep checks for the greppable ones

Adopt Before You Build

In an app that already has screens, the first move is detection, not construction. Before writing any token file:

1. Look for a declared system. Check package.json for a styling library - NativeWind/Tailwind, Tamagui, Restyle, Unistyles, styled-components. Then look for a token file: theme.ts, src/theme/, constants/theme.ts, or constants/Colors.ts (the create-expo-app default).
2. If one exists, it is the source of truth. Extend it in its own idiom - its names, its scale, its storage format. Audit drift against that system, not against the examples below.
3. If only de facto values exist - the same greys and paddings repeated across screens, no theme file - there is no system yet. Those values are the input to the scales, not the authority: derive tokens from the most frequent ones, snapped to the 4-point grid (references/audit.md §5).
4. Never introduce a second system beside an existing one. A fresh src/theme/ next to a Tamagui config is design-system drift, not adoption.

Only when nothing exists do the defaults below apply as written.

The Theme

In an app without an existing system, all design tokens live under src/theme/. In a project without a src/ folder (the default create-expo-app template has app/, components/, and constants/ at the root), use the equivalent top-level location - typically theme/ or the existing constants/ - and keep the same file layout. Start small and split by token class as it grows:

src/theme/
  colors.ts       # see expo-native-ui "Colors" for the palette pattern
  spacing.ts
  typography.ts

A brand-new app can begin with a single src/theme.ts holding all of the objects below, then promote it to the folder form once any one class needs its own file (same promotion rule as components). Either way there is exactly one theme entry point - never two competing token files.

Rules that make a theme worth having:

• Every repeated visual value is a token. A literal that appears twice belongs in the theme.
• Components import tokens; screens import components. A screen file that imports spacing for layout padding is fine; a screen file redefining a button color is drift.
• Never hardcode hex colors, font sizes, or spacing multiples outside src/theme/. One-off values that are genuinely local (an icon's 17px optical nudge) may stay inline - with a comment saying why.

Colors

Build the palette from platform semantic colors: Color from expo-router wrapped in Platform.select, centralized in theme/colors.ts. Semantic colors resolve on-device and adapt to light/dark automatically - prefer them for backgrounds, labels, and separators. (expo-native-ui "Colors" covers the full palette and rationale; the minimal version is:)

Add brand colors as explicit light/dark pairs only when the brand requires values the platform doesn't provide:

Keep the brand set tiny (accent, accentContrast, maybe a tint per feature). Everything else stays semantic.

Static-safe vs hook-only. The two patterns above have different reach - keep the boundary explicit:

• Semantic/platform colors (colors above) are static-safe: they resolve on-device, so plain token files like theme/typography.ts can import them at module scope.
• Brand light/dark pairs are hook-only: useBrandColors() reads the color scheme at render time, so brand colors can only be applied inside components. A static token file cannot call the hook.
• Never mix the two in one file. If a static style (a type ramp step, a variants object) needs the brand accent, either apply the brand color in the component at render time, or wrap the pair in a static dynamic color (DynamicColorIOS on iOS) so it becomes static-safe.

Spacing

One scale, based on a 4-point grid. Name steps by size, not by use:

• Use gap with spacing tokens for layout rhythm (expo-native-ui prefers gap over margin).
• Screen edge padding is spacing.md unless the design says otherwise - pick one and keep it.
• If a layout needs a value between steps, use the nearest step. The grid is the point.
• If the same in-between multiple of 4 keeps recurring (12 and 20 are common), add it to the scale as a named step instead of scattering literals. The audit whitelist must then include it too.

Typography

Define named text styles, not raw font sizes. Mirror the platform ramp (Apple text styles) so sizes feel native:

If the project bundles static font files (one file per weight, loaded with expo-font or the config plugin), set weight via fontFamily names instead and omit fontWeight - otherwise iOS synthesizes the weight or falls back to the system font:

Expose them through one component so screens never touch fontSize:

Screen titles still come from the navigation stack header (expo-native-ui rule), so largeTitle is mostly for non-stack contexts.

Dynamic Type. Text scales with the user's system text-size setting (allowFontScaling is on by default). Use padding or minHeight around text so rows can grow, and check large accessibility text sizes. Let labels wrap or reflow before considering a per-element maxFontSizeMultiplier for constrained chrome; dense rows alone are not a reason to cap readable text. Never disable scaling app-wide with allowFontScaling={false}.

Radius

Pair every non-capsule radius with borderCurve: "continuous" (per expo-native-ui).

Shadows

Shadows are boxShadow strings (never legacy shadow/elevation props - see expo-native-ui). Two or three elevation levels are enough:

Motion

Durations and shared spring/easing configs, so animations across the app feel related:

Reanimated caveat: don't pass Color/PlatformColor token values into Reanimated styles - use static colors there (see expo-native-ui).

Reusable Components

The theme controls values; components control structure. Shared primitives live in src/components/ (see expo-project-structure).

The component contract

Every design-system primitive defines, explicitly:

• Variants - visual intent: primary, secondary, ghost, destructive. Add a variant only when a real screen needs it.
• Sizes - sm, md, lg. Default md. Sizes map to spacing/typography tokens, never to fresh numbers.
• States - default, pressed (not hover - this is touch), disabled, loading. Handle pressed with a Pressable style function; never leave a tappable element without pressed feedback.
• Style override - accept a style prop and merge it last, so callers can adjust layout (margins, flex) without forking the component. Callers may override layout, not identity - a caller changing a button's colors is a signal the variant set is missing something.
• Accessibility - custom interactive primitives expose their role and disabled/busy/selected state as applicable. Text children can supply the label; icon-only controls and buttons that replace text with a spinner need an explicit label that remains available while loading. Verify labels on native controls too.

Composition over configuration

When a component's props start describing content (leftIcon, subtitle, footerText, badgeCount), stop adding props and accept children instead. A Card that renders children with token padding outlives any Card with twelve content props. Reserve props for the contract above: variant, size, state, style.

When to extract - and when not to

Promote a view into src/components/ when all of these hold:

1. It appears (or is about to appear) in two or more screens. Until then it stays colocated in screens/<name>/ (see expo-project-structure).
2. It has a nameable role ("Card", "EmptyState", "Badge") - not "the thing on the profile screen".
3. Its API is smaller than its implementation. If the props would just re-expose every internal style, it isn't a reusable component yet - it's a screen fragment.

Promotion path: inline JSX → component in screens/<name>/ → src/components/. Move one step at a time, when the trigger fires - never speculatively. Wrong abstractions cost more than duplication; a second copy of a view is cheaper than a primitive with a bad API.

Do not wrap platform components that already carry the design language (Switch, DateTimePicker, stack headers, @expo/ui views) just to route them through the system. Native styling is the design system for those.

Where Decisions Live

Self-Critique Pass

After building or changing a screen, screenshot it and check it against these principles (from Expo's design-principles guide). Each one maps to a system fix, not a local tweak:

• Hierarchy / contrast - is the most important element obviously first? Fix with type ramp steps, not ad-hoc font sizes.
• Proximity / white space - do related items sit closer than unrelated ones? Fix with gap + spacing tokens.
• Repetition / unity - do all corners, shadows, and accents match? If not, a value escaped the theme - move it in.
• Alignment - do edges share axes? Fix with consistent screen edge padding.

Recheck the rendered result after fixing a value; moving it into the theme does not itself fix the layout. If the same defect recurs across screens, fix the shared token or component. Also run the primary-task and content checks in expo-native-ui's Behavior section; screenshots alone cannot verify interaction.

Named Failures: Native Slop

Use these names to recognize common mistakes when building and reviewing:

• The Web Modal - a custom centered dialog used for composing or picking. Prefer a native sheet (formSheet, @expo/ui BottomSheet) or menu; native confirmation alerts remain appropriate for consequential actions.
• Everything's a Card - every row and section in its own white rounded shadowed box. Use grouped lists; group with background and hairlines, not borders.
• Emoji Iconography - 🔥 ⚙️ ✨ as tab or button icons. SF Symbols on iOS, Material icons on Android.
• The Purple-Gradient Hero - a decorative gradient intro pushing the task below the fold. Lead task screens with useful content; retain a hero when it serves the requested experience.
• The Spinner Blink - a full-screen spinner between every state, or "No items yet" flashing during the first load. Every screen has four states (see expo-data-fetching).

Treat visual tells as review prompts, not blanket bans on cards, fonts, or branding. Fix the observable problem and respect the user's brief and existing design system. The full list of 20 and candidate grep checks are in ./references/native-slop.md; use them to explain the problem and replacement when reviewing a screen.

Auditing an Existing App

To measure drift in an app that already has screens - hardcoded hex values, arbitrary spacing, inconsistent component APIs - follow ./references/audit.md. It contains grep-based checks, a scoring rubric, an incremental adoption order for fixing a drifted app, and templates for documenting existing components and proposing new ones.

Submitting Feedback

If you encounter errors, misleading or outdated information in this skill, report it so Expo can improve:

Only submit when you have something specific and actionable to report. Include as much relevant context as possible. If an AI agent repeatedly failed or the user had to take over an Expo task, load the expo-skill-feedback skill and follow its eval-candidate flow instead of reusing the command above.