Siddhartha's Flutter app conventions and architecture patterns. Use when writing Flutter/Dart app code, scaffolding a new Flutter feature or whole project, designing state management, networking, offline-first sync, or dependency injection, or making architecture decisions in a Flutter project. Also use when reviewing Flutter PRs, structuring packages, or when someone asks about Flutter project conventions. Ensures consistency across Flutter projects and shared work.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Siddhartha's Flutter app conventions and architecture patterns. Use when writing Flutter/Dart app code, scaffolding a new Flutter feature or whole project, designing state management, networking, offline-first sync, or dependency injection, or making architecture decisions in a Flutter project. Also use when reviewing Flutter PRs, structuring packages, or when someone asks about Flutter project conventions. Ensures consistency across Flutter projects and shared work.
Siddhartha's Flutter Style Guide
This guide encodes the architecture, conventions, and preferred stack for Siddhartha's Flutter apps. Follow these patterns for all Flutter work and shared projects. The goal is a single, repeatable shape so that building a new feature is mechanical and an agent can clone the pattern without re-deciding architecture each time.
This is a Flutter-native guide. It does not borrow backend layering — Flutter is a reactive, declarative widget tree, and the patterns here are built for that, not translated from server code.
⚠️ Verify current versions before scaffolding
The Flutter ecosystem moves fast and several of these libraries have had breaking API changes. Before writing or scaffolding code, confirm the current syntax — do not trust the snippets here as eternal truth.
Prefer context7: call resolve-library-id then query-docs for freezed, riverpod / riverpod_generator, retrofit, drift, go_router_builder.
Fall back to web search on pub.dev for the latest version and migration notes.
The code in this skill's references/ reflects freezed 3.x (abstract/sealed class keywords) and Riverpod 3 (unified type) as of authoring. Confirm these still hold, and adjust if a newer major has shipped.
Ref
Preferred stack
Concern
Choice
Why
State + DI
Riverpod 3 + riverpod_generator (both)
One library covers reactive state and the dependency graph; provider overrides are the test seam; the graph is analyzer-checked, unlike a runtime service locator
Local widget state
flutter_hooks + hooks_riverpod
Ephemeral controllers / focus / animation without StatefulWidget boilerplate; composes with Riverpod via HookConsumerWidget (same author)
Models / JSON
freezed + json_serializable
Immutable classes, copyWith, value equality, sealed unions, JSON — one annotation per type
Networking
dio + retrofit
dio for interceptors/retries/cancellation; retrofit generates the typed client from an annotated abstract class
Errors
sealed Failure + dio interceptor
Mapped once, thrown, caught by AsyncValue; exhaustive typed handling without a second error channel
Routing
go_router + go_router_builder (typed)
Compile-safe navigation; a typed route class is a clean per-feature template slot
Local / offline
drift (opt-in)
Reactive SQLite as local source of truth — only when the app is offline-first (see below)
Testing
mocktail + in-memory fakes
Stateful fakes for repositories, mocktail for leaf boundaries
Animation
flutter_animate (+ the flutter-animations skill if installed)
Easy-mode common animations; the companion skill covers explicit/physics/Hero
Dependency philosophy
Judge a dependency by how much it entangles with the others (n²), not by the raw count (n+1). A focused, well-aligned dep with one clean integration seam is close to free; a dep that overlaps responsibilities with something you already run is expensive no matter how small, because it multiplies the "which one owns this?" decision across the whole codebase.
A good (n+1) dep is self-contained, has one intentional integration point, and doesn't fight the rest of the stack. flutter_hooks qualifies — it touches only Riverpod, through the first-class HookConsumerWidget.
A bad (n²) dep competes with a tool you already have. That — not "fewer deps" — is the real reason these are out:
get_it — a second DI system racing Riverpod's provider graph.
bloc — a second state system racing notifiers.
fpdart by default — a second error channel racing AsyncValue (worth it only locally inside a chaining-heavy repository; see the error escape hatch).
GetX — competes with routing, DI, and state at once. The worst kind of n².
When weighing any new package, ask: does it have one clean seam, or does it overlap with two things I already have?
Architecture: MVVM, feature-first
Every feature is a self-contained folder with three layers. Data flows up as state; events flow down as commands.
Repository — the source of truth for a domain concept. Calls data-sources, maps DTO → domain, throws Failure.
Data-sources — the retrofit API client (remote) and, when offline-first, the drift database (local).
There is no Clean Architecture use-case layer — single-action interactor classes are almost always one-line pass-throughs. The ViewModel calls the repository directly.
Keep wire data, domain data, and view data separate. The wire shape must never reach a widget.
DTO (data/) — matches the JSON exactly. @JsonSerializable (or a freezed class with fromJson).
Domain model (domain/) — the app's canonical entity, a @freezed class. The repository maps DTO → domain.
View state (presentation/) — what the ViewModel exposes to the View.
Rules:
The DTO ↔ domain wall is always enforced. A backend field rename stops at the repository; it never ripples to the UI.
A dedicated view-state class is minted only when the screen composes or derives — merges multiple entities, adds UI-only fields (a filter, a selection), or computes values. For a straight list, AsyncValue<List<Expense>>is the view state; a wrapper class would be empty ceremony.
A sealed class Failure lives in core/network/failure.dart:
sealed class Failure implements Exception {
const Failure(this.message);
final String message;
}
class NetworkFailure extends Failure { const NetworkFailure([super.m = 'No connection']); }
class ServerFailure extends Failure { const ServerFailure(super.m, this.statusCode); final int statusCode; }
class UnauthorizedFailure extends Failure { const UnauthorizedFailure([super.m = 'Session expired']); }
class ValidationFailure extends Failure { const ValidationFailure(super.m, [this.fields = const {}]); final Map<String, String> fields; }
class UnknownFailure extends Failure { const UnknownFailure([super.m = 'Something went wrong']); }
Errors are converted to a thrown Failure by two small pieces in core/network/:
// classifier — DioException → Failure
Failure mapDioException(DioException e) => switch (e.type) {
DioExceptionType.connectionError ||
DioExceptionType.connectionTimeout ||
DioExceptionType.receiveTimeout => const NetworkFailure(),
DioExceptionType.badResponse => switch (e.response?.statusCode) {
401 => const UnauthorizedFailure(),
422 => ValidationFailure(e.response?.data?['error']?.toString() ?? 'Invalid'),
final c? when c >= 500 => ServerFailure('Server error', c),
_ => const UnknownFailure(),
},
_ => const UnknownFailure(),
};
// the one place DioException is unwrapped — repositories call through this
Future<T> guard<T>(Future<T> Function() body) async {
try {
return await body();
} on DioException catch (e) {
throw mapDioException(e);
}
}
Repositories call through guard(). That is the single place DioException becomes a Failure. (A dio interceptor cannot do this alone: ErrorInterceptorHandler.reject only accepts a DioException, so the type that escapes an await is always DioException, never your Failure. The conversion must happen in a Dart try/catch. Use interceptors for auth/logging/retry, not for error typing.)
The ViewModel does not try/catch — the repository already throws Failure, and Riverpod's AsyncValue captures it.
The View handles errors by switching the Failure in the error branch (exhaustive, compiler-checked).
Escape hatch: a feature with genuine multi-step domain logic (validate → transform → persist → sync) may use Either/Result (fpdart) locally inside its repository, then collapse it to a throw at the ViewModel boundary (result.fold((f) => throw f, (d) => d)) so it re-enters the single AsyncValue pipe. fpdart is not a default dependency — pull it in only for the feature that needs chaining.
Offline-first (opt-in)
Most apps are not offline-first, and the offline machinery is real, ongoing complexity. So:
Before scaffolding any local store, mutation queue, or sync engine, ASK the developer: "Is this app offline-first?" Only build it if the answer is yes.
When it is yes, follow references/offline-first.md: drift becomes the local source of truth (the UI reads drift, writes hit drift instantly), a mutation-outbox table queues changes, and a pluggable sync engine drains the queue against a delta endpoint with last-write-wins + surface-conflicts resolution. The default online slice does not load any of this.
UI layer
Views are dumb. Logic lives in the ViewModel.
Widgets are HookConsumerWidget (or ConsumerWidget when no hooks are needed). They watch providers and render — nothing else.
State ownership — the bright line that keeps hooks at n+1: hooks hold ephemeral, widget-local state that dies with the widget (useTextEditingController, useFocusNode, useAnimationController, useState for a local toggle). Riverpod holds anything shared or longer-lived than the widget. If state must survive a rebuild-from-scratch or be read elsewhere, it's a provider, never a hook.
Split a big build() into small, named, const leaf widgets in widgets/ — never one 300-line build().
A widget that doesn't depend on a BuildContext value should be const.
Scope rebuilds with ref.watch(provider.select(...)) so a widget rebuilds only when the slice it uses changes — don't watch a whole object to read one field.
Use ListView.builder / SliverList for any list of unknown length — lazy by default, builds only visible items.
Add Keys where widget identity matters (reorderable / filtered lists).
Platform-adaptive look by default — Cupertino-flavoured on iOS, Material 3 on Android. All platform branching lives in one adaptive layer (core/ui/: AppButton, AppScaffold, …); feature code never sees Platform.isIOS.
Theming: Material 3 ThemeData + ThemeExtension for brand tokens. No literal colors / spacing / radii in a build() — pull from the theme or a token.
iOS Liquid Glass is not in core Flutter (design updates paused in 2025); only young community packages bridge it. Opt-in, and isolated behind core/ui/ so it's swappable later.
Full UI treatment — adaptive-layer code, the ThemeExtension token example, the Liquid Glass libraries, and widget organization — is in references/ui-layer.md.
Codegen workflow
Run dart run build_runner watch -d during development.
Generated files (*.g.dart, *.freezed.dart) are gitignored. CI runs dart run build_runner build --delete-conflicting-outputs before analyze/test.
A fresh clone shows errors until the first generate — this is expected; running build_runner is step one (the project-setup reference makes this explicit).
Naming
Files: snake_case.dart. One file per layer-piece per feature.
Types: PascalCase. ViewModels named for the screen (ExpenseList, not ExpenseListViewModel — the @riverpod class ExpenseList generates expenseListProvider).
Providers: <thing>Provider (generated automatically from the annotated name).
Starting a new project
For zero → first running feature (pubspec with the stack, folder skeleton, analysis_options.yaml, build_runner, run commands), follow references/project-setup.md.
Testing
In-memory fakes for repositories, mocktail at the boundaries, ProviderScope/ProviderContainer overrides as the seam. Full patterns in references/testing.md.
Anti-patterns
get_it for DI — Riverpod's provider graph is the DI, and it's analyzer-checked.
bloc — Riverpod is the state solution here; don't mix two.
fpdart by default — only inside a chaining-heavy repository (see escape hatch).
GetX — magic, global state, poor testability. Never.
Letting a DTO / JSON map reach a widget — always map to a domain model first.
Wrapping Either inside AsyncValue — two competing error channels; throw instead.
Business logic in a build() method or a widget — it belongs in the ViewModel/repository.
A 300-line build() — split into const leaf widgets.
Building offline/sync machinery without asking whether the app is even offline-first.
