agentleFS
zakariafzakariaf/Flutter-Skills/skills/flutter-performance/SKILL.md

flutter-performance

Skill32 starsChanged 4 months agoMITView on GitHub

Enforces Flutter runtime performance — const subtrees, minimal rebuild scope via ref.watch(select), lazy ListView/GridView builders and slivers, sized image decode (cacheWidth/ResizeImage), heavy work off the UI isolate via compute/Isolate, surgical RepaintBoundary, dispose everything, and measurement in profile mode on a floor device. Use when optimizing UI, diagnosing jank or dropped frames, tuning long lists or images, reviewing rebuild/repaint scope, or when the task mentions const, select, ListView.builder, cacheWidth, compute, RepaintBoundary, AnimatedBuilder, DevTools, raster thread, or 60/120fps.

3 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Shrink the rebuild scope to the smallest changing widget: Never ref.watch(provider) for a whole state object when one field changed — watch …

    line 13
  2. Distinctive:

    Use "Track widget rebuilds" to find widgets rebuilding too often.

    line 86
  3. Very specific:

    ref.watch(provider) for a whole object in a leaf: when one field changed — use .select.

    line 94

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Flutter Performance

Hold a steady 60 fps (<=16 ms/frame; <=8 ms on 120 Hz) on a floor device (low-end Android / older iPhone). Performance is a property you measure in profile mode, not one you assert. It comes from rebuilding less, painting cheaply, and keeping both the UI thread and the raster thread free.

Non-negotiable rules

  1. const everything legal. A const subtree is skipped on rebuild — the framework short-circuits it by identity. Keep prefer_const_constructors on as an error; a non-const literal that could be const is a lint failure, not a preference.
  2. Shrink the rebuild scope to the smallest changing widget. Never ref.watch(provider) for a whole state object when one field changed — watch provider.select((s) => s.field). A HUD counter tick must never rebuild a heavy sibling. Rebuild scope is the single biggest lever on build-thread cost.
  3. Lazy everything long. ListView.builder / GridView.builder / slivers for variable or unbounded content. ListView(children: items.map(...).toList()) builds every off-screen row up front — refuse it for anything not tiny and fixed.
  4. No expensive work in build(). No jsonDecode, sort, regex, DateTime math, file/network, or large allocation — build() may run every frame. Compute in the Notifier/ViewModel once and cache the result in immutable state.
  5. Heavy CPU off the UI isolate. Parse large payloads, process images, crunch numbers in compute() or a spawned Isolate. Blocking the UI thread is guaranteed jank; the raster thread cannot save you.
  6. Size image decode to the display slot. cacheWidth/cacheHeight or ResizeImage, correct asset resolutions, precacheImage for above-the-fold art. Decoding a 4000 px source into a 100 px box spikes memory and OOMs cheap phones.
  7. Prefer the cheapest widget that does the job. ColoredBox/DecoratedBox over Container, SizedBox for spacing, FadeTransition/AnimatedOpacity over the Opacity widget in hot paths. Avoid ClipPath/saveLayer inside scrolling lists — they force an offscreen buffer on the raster thread.
  8. RepaintBoundary around costly, independently-repainting subtrees (an animation, a chart, a live indicator) so its repaint does not re-raster its neighbours. Do not sprinkle boundaries everywhere — each is a compositor layer that costs memory.
  9. Pass expensive children through child:. AnimatedBuilder/ValueListenableBuilder/StreamBuilder rebuild only the builder; hand the unchanging subtree in via child: so it is built once, not per frame.
  10. Dispose to avoid leaks. AnimationControllers, gesture recognizers, stream subscriptions, timers, TextEditingControllers, caches — released in dispose() (widgets) or ref.onDispose (providers). autoDispose scoped session state so it does not survive the screen.
  11. Measure in PROFILE mode on a real floor device with DevTools. Debug timings are meaningless (JIT, asserts). Watch the raster thread as well as the UI thread — saveLayer/blur/clip cost shows there, not in Dart. Never ship a frame-budget number that was asserted, not measured.

Narrow the rebuild scope

Watch the single field that changes, not the whole state object:

// Rebuilds only when the count changes, not on every OrderState mutation.
class OrderBadge extends ConsumerWidget {
  const OrderBadge({super.key});

  @override
  Widget build(BuildContext context, WidgetRef ref) {
    final count = ref.watch(orderProvider.select((s) => s.items.length));
    return Text('$count');
  }
}

.select runs the selector every publish but rebuilds only when the selected value's == changes, so its input must be an immutable value with real equality. See state-management-riverpod for the watch/read/listen split and family + autoDispose.

Keep expensive children out of the animation loop

// The card is built once; only the transform recomputes each tick.
AnimatedBuilder(
  animation: _controller,
  child: const ProductCard(),
  builder: (_, child) =>
      Transform.scale(scale: _controller.value, child: child),
);

Lazy lists and sized images

// Lazy — only visible rows are built.
ListView.builder(
  itemCount: items.length,
  itemBuilder: (context, i) => ItemTile(item: items[i]),
);

// Decode to the slot, not the source resolution.
Image.asset('assets/hero.png', cacheWidth: 240);

Heavy work off the UI isolate

// parseItems is a top-level or static function (isolate entry-point rules).
final items = await compute(parseItems, jsonString);

Custom painting

For a CustomPainter, the discipline is zero-allocation paint(), a shouldRepaint that is one value comparison over an immutable scene value, and a repaint: Listenable (the AnimationController) instead of setState in a ticker. That treatment lives in custom-canvas-and-gestures; the one performance-critical rule to carry here: never allocate Paint/Path/Gradient inside paint() — hold them as painter fields and mutate cheap properties.

Profiling workflow

  1. flutter run --profile on a physical floor device.
  2. DevTools -> Performance; record while reproducing the jank.
  3. For each over-budget frame, check whether the time is in Build, Layout, or Raster.
    • Build-heavy -> narrow rebuild scope with .select, add const, move work out of build().
    • Raster-heavy -> reduce Opacity/clips/saveLayer, add a RepaintBoundary, simplify shaders.
  4. Use "Track widget rebuilds" to find widgets rebuilding too often.
  5. Fix one bottleneck, re-measure. Never optimize without a before/after measurement.

Anti-patterns

  • Judging performance in debug mode — JIT and asserts make timings fiction. Profile mode, real device, always.
  • Premature optimization with no profile data — adds complexity, hides nothing; measure first.
  • setState() high in the tree rebuilding a whole screen for a tiny change — scope the state down.
  • ref.watch(provider) for a whole object in a leaf when one field changed — use .select.
  • ListView(children: [...].toList()) for long/unbounded content — builds every off-screen row.
  • jsonDecode/sort/DateTime math/regex in build() — it may run every frame; precompute in the ViewModel.
  • Opacity widget for fades in scrolling lists — forces saveLayer; use FadeTransition or fade via color.
  • Decoding full-resolution images into small widgets — memory spikes and OOM on cheap phones.
  • Parsing large payloads / heavy loops on the UI isolate — frozen frames; use compute.
  • RepaintBoundary on everything — layer-memory bloat; use it surgically around independently-repainting subtrees.
  • AnimatedBuilder/StreamBuilder rebuilding an expensive child instead of passing it via child:.
  • Rebuilding on every keystroke without debounce, or unbounded caches/listeners that leak.
  • Allocating Paint/Path inside paint() — per-frame garbage; hold them as fields.

Definition of done

  • const applied everywhere legal; prefer_const_constructors clean.
  • Rebuilds scoped with .select; a tick in one widget never rebuilds a heavy sibling.
  • Long/variable lists are lazy (.builder/slivers).
  • No I/O, parsing, or heavy compute in build(); big work runs via compute/Isolate.
  • Images sized to display (cacheWidth/ResizeImage); above-the-fold precached.
  • Cheapest suitable widgets/effects chosen; RepaintBoundary used surgically; expensive children passed via child:.
  • Any CustomPainter allocates nothing in paint(); shouldRepaint is one value compare.
  • Controllers, recognizers, subscriptions, timers disposed; scoped session state autoDisposed; no leaks.
  • Measured in profile mode on a floor device via DevTools — UI and raster threads under budget; no frames over 16 ms; claim backed by a recording.

Related skills

  • state-management-riverpod — the watch/read/listen split, .select, family + autoDispose, immutable state that makes .select correct.
  • widget-composition — small const Widget classes over _buildX methods, dispose discipline, cheapest-widget choices.
  • custom-canvas-and-gestures — zero-allocation paint(), shouldRepaint as one value compare, repaint: Listenable.
  • testing-strategy — clock-injected pure core so heavy compute is testable off the UI isolate.

References

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.