dpearson2699/swiftui-performance
Profile, diagnose, and remediate SwiftUI runtime performance using code review, Instruments, and repeatable measurements. Use when a SwiftUI screen renders slowly, scrolling or animations hitch, view bodies update excessively, list identity churns, layout work spikes, or broad Observation dependencies raise CPU cost. Covers evidence-based triage, SwiftUI Instruments lanes, lazy-container guardrails, state lifetime, and before/after verification.
npx skills add https://github.com/dpearson2699/swift-ios-skills --skill swiftui-performance
Audit SwiftUI view performance from a reproducible symptom to measured
remediation. Route animation design to swiftui-animation, production telemetry
to metrickit, ownership/leak analysis to ios-memgraph-analysis, navigation
behavior to swiftui-navigation, state architecture to swiftui-patterns, and
layout construction to swiftui-layout-components.
device, OS, and build configuration.
broad refactor.
Use this triage list for both code and trace analysis:
bodyMap each suspect from the triage list to exact code. Report likely causes with
code references, but label them code-backed hypotheses until a trace confirms
cost. Propose a minimal repro or measurement when evidence is missing.
Use the SwiftUI Instruments template on a Release build and real device when
possible. Reproduce the exact interaction, capturing SwiftUI lanes, Time
Profiler, and Hangs/Hitches as relevant. Ask for the trace or screenshots of the
lanes and call tree.
Apply the same triage list to trace evidence. Correlate long or frequent SwiftUI
updates with the Time Profiler call tree and the reproduced interaction. Separate
trace-backed findings from code-backed hypotheses and name the next measurement
that would resolve remaining uncertainty.
Apply targeted fixes:
@State/@Observable closer to leaf views).ForEach and lists.body into model-layer precomputation, an explicit derivedvalue updated when its inputs change, a memoized helper, or background processing.
Use @State only when the view owns both the value and its update lifecycle; it is
not a generic cache for arbitrary computation.
equatable() only when equality is cheaper than recomputing the subtree andthe compared inputs have stable value semantics.
| Smell | Evidence to seek | Targeted fix |
|---|---|---|
| Formatter, sort, filter, or decode in body | Long/frequent body updates with matching call-tree cost | Recompute when inputs change; downsample/decode off the main actor |
| UUID() or unstable id: \.self | Recreated rows, lost state, excess updates | Use stable model identity |
| Root if/else swaps | State reset or update spikes when toggled | Localize conditional content/modifiers when semantics allow |
| Broad model reads | Many unrelated views update together | Pass narrow values or move reads into focused child views |
| Geometry writes during layout | Repeating layout/update cycle | Threshold changes or replace the feedback path with stable layout |
Ask the user to re-run the same capture and compare with baseline metrics.
Summarize the delta (CPU, frame drops, memory peak) if provided.
Provide:
Use the SwiftUI template in Instruments (Cmd+I to profile). Current SwiftUI lanes include Update Groups, Long View Body Updates, Long Representable Updates / Representable Updates, Other Long Updates / Other Updates, and the Cause & Effect Graph. Correlate those with Time Profiler and Hangs/Hitches.
Add Self._printChanges() in debug builds to log which property triggered a view update:
var body: some View {
#if DEBUG
let _ = Self._printChanges() // "MyView: @self, _count changed."
#endif
Text("Count: \(count)")
}
See references/optimizing-swiftui-performance-instruments.md for the full profiling workflow.
Identity controls view lifetime and state. Use stable model IDs in repeated
content and reserve .id(_:) changes for intentional resets. Prefer
@ViewBuilder or generic composition over AnyView in profiled hot rows. Treat
root conditional branches as suspects—not automatic defects—when evidence shows
state churn or expensive recreation.
Text(title)
.foregroundStyle(isHighlighted ? .yellow : .primary)
ForEach(items) { item in
Row(item: item).id(item.stableID)
}
Use lazy containers when profiling shows eager construction, layout, or update
work is material; there is no universal item-count threshold. Route grid/list
construction choices to swiftui-layout-components.
Guardrails:
onAppear because of prefetching. Do not make onAppear the only setup point for data a row needs to render.onAppear and onDisappear as visibility signals, not lifetime guarantees.ForEach; avoid if branches that make each element produce zero or one row.ForEach element to a constant number of top-level subviews. Wrap row contents in a stable container if needed. Use -LogForEachSlowPath YES while debugging list/table slow paths.Layout before feeding geometry changes back into row state.Observation tracks properties read during view evaluation. Reduce fan-out by
passing narrow derived values or moving reads into focused child views.
// Split reads into child views so each tracks only what it renders.
struct ProfileView: View {
let model: ProfileModel
var body: some View {
VStack {
NameRow(model: model) // only tracks name
EmailRow(model: model) // only tracks email
AvatarView(model: model) // only tracks avatar
SettingsForm(model: model) // only tracks settings
}
}
}
Cheap computed values can remain derived; expensive transformations need an
explicit owner, input set, and refresh trigger. Do not add view models as a
performance ritual—measure first and route general state design to
swiftui-patterns.
@Observable models into focused ones, or use computed properties/closures to narrow observation scope..onGeometryChange (iOS 16+) with thresholds.DateFormatter() or NumberFormatter() inside body. These are expensive to create. Make them static or move them outside the view.Equatable, then use .animation(_:value:) for simple value-bound changes or .animation(_:body:) for narrower modifier-scoped implicit animation.List without identifiers. Use id: or make items Identifiable so SwiftUI can diff efficiently instead of rebuilding the entire list.@State wrapper objects. Wrapping a simple value type in a class for @State defeats value semantics. Use plain @State with structs.MainActor with synchronous I/O. File reads, JSON parsing of large payloads, and image decoding should happen off the main actor. Prefer nonisolated async helpers or dedicated actors; reserve Task.detached for cases where you intentionally break actor inheritance and handle cancellation yourself.DateFormatter/NumberFormatter allocations inside bodyIdentifiable items or explicit id:@Observable models expose only the properties views actually readMainActor (image processing, parsing)onAppear as the only setup point.animation(_:value:) for value-bound changes or .animation(_:body:) for narrower modifier scope@State is not used as an unspecified cache; every derived value has an explicit owner and refresh triggerequatable() is used only when comparison is cheaper than recomputation and inputs have stable value semantics@Observable view models are @MainActor-isolated; types crossing concurrency boundaries are SendableTake dpearson2699/swiftui-performance from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.