Frontend Architecture

Purpose

Structure a frontend so that a new engineer can find the code for a feature in under a minute, and so that changing one feature does not require touching five others.

When to Use

• Starting a new frontend application.
• A codebase where features are scattered across components/, utils/, hooks/, and types/.
• Deciding where state belongs and what owns data fetching.
• Setting up a monorepo or splitting a large application.

Capabilities

• Feature-based organization and module boundaries.
• State layering: server cache, global client state, feature state, component state.
• Data-access layer design and API client structure.
• Build configuration, code splitting, and bundle budgets.
• Monorepo structure and shared package design.

Inputs

• The application's feature set and team structure.
• Current pain: where changes ripple, what is hard to find.
• Build tooling and deployment target.

Outputs

• A folder structure organized by feature, not by file type.
• Explicit rules for what may import what.
• Bundle budgets enforced in CI.

Workflow

1. Organize by feature, not by kind — features/checkout/ containing its components, hooks, api, and types beats components/, hooks/, api/ each containing a slice of every feature.
2. Draw the import rules — Features may import from shared/. Features may not import from each other; if two need the same thing, it moves to shared/. Enforce this with a lint rule, not a convention document.
3. Layer the state deliberately — Server data in a query cache. Genuinely global client state (theme, session) in one store. Everything else stays local. Most "global state" is server state in disguise.
4. Centralize the API client — One place that knows about base URLs, auth headers, error mapping, and retries. Feature code calls typed functions, not fetch.
5. Budget the bundle — Set a size limit per route and fail the build when it is exceeded. Bundle size regresses one dependency at a time.

Best Practices

• A utils/ folder is where code goes to be forgotten. If a function belongs to a feature, keep it in the feature.
• Barrel files (index.ts re-exporting everything) defeat tree-shaking and create import cycles. Import from the source module.
• Route-level code splitting is nearly free and pays for itself immediately. Component-level splitting rarely does.
• Do not put server data in a global store. It has staleness, refetch, and error semantics that a store does not model — you will rebuild a query library, badly.
• A shared component library inside the app is fine. Extracting it into a package before a second consumer exists is premature.
• Keep the dependency count low. Every dependency is bundle weight, a supply-chain risk, and a future migration.

Examples

Feature-based structure with enforced boundaries:

src/
  app/                    # routing, providers, global layout
  features/
    checkout/
      components/         # only used by checkout
      api/                # checkout endpoints, typed
      model/              # checkout state and domain types
      index.ts            # the feature's public surface
    orders/
    account/
  shared/
    ui/                   # design-system primitives
    api/                  # http client, auth, error mapping
    lib/                  # genuinely cross-cutting helpers

// eslint.config.js — the boundary is enforced, not merely documented.
{
  rules: {
    "import/no-restricted-paths": ["error", {
      zones: [{
        target: "./src/features/*",
        from: "./src/features/*",
        message: "Features must not import each other. Move shared code to src/shared.",
      }],
    }],
  },
}

A bundle budget that fails the build:

{
  "bundlesize": [
    { "path": "dist/assets/index-*.js", "maxSize": "180 kB", "compression": "brotli" },
    { "path": "dist/assets/checkout-*.js", "maxSize": "90 kB", "compression": "brotli" }
  ]
}

Notes

• The single most effective architectural rule in a frontend codebase is "features may not import each other". It is trivially enforceable and prevents the coupling that makes large frontends unchangeable.
• Analyze the bundle before optimizing it. vite-bundle-visualizer or source-map-explorer will usually show one date library or one icon set accounting for a third of the payload.
• Monorepos solve a versioning problem between packages. If you have one application, a monorepo is overhead with no corresponding benefit.