Swiftui Expert Skill

SwiftUI Expert Skill

Operating Rules

• Consult references/latest-apis.md at the start of every task to avoid deprecated APIs
• Prefer native SwiftUI APIs over UIKit/AppKit bridging unless bridging is necessary
• Focus on correctness and performance; do not enforce specific architectures (MVVM, VIPER, etc.)
• Encourage separating business logic from views for testability without mandating how
• Follow Apple's Human Interface Guidelines and API design patterns
• Only adopt Liquid Glass when explicitly requested by the user (see references/liquid-glass.md)
• Present performance optimizations as suggestions, not requirements
• Use #available gating with sensible fallbacks for version-specific APIs

Task Workflow

Review existing SwiftUI code

• Read the code under review and identify which topics apply
• Flag deprecated APIs (compare against references/latest-apis.md)
• Run the Topic Router below for each relevant topic
• Validate #available gating and fallback paths for iOS 26+ features

Improve existing SwiftUI code

• Audit current implementation against the Topic Router topics
• Replace deprecated APIs with modern equivalents from references/latest-apis.md
• Refactor hot paths to reduce unnecessary state updates
• Extract complex view bodies into separate subviews
• Suggest image downsampling when UIImage(data:) is encountered (optional optimization, see references/image-optimization.md)

Implement new SwiftUI feature

• Design data flow first: identify owned vs injected state
• Structure views for optimal diffing (extract subviews early)
• Apply correct animation patterns (implicit vs explicit, transitions)
• Use Button for all tappable elements; add accessibility grouping and labels
• Gate version-specific APIs with #available and provide fallbacks

Record a new Instruments trace

Trigger when the user asks to "record a trace", "profile the app", "capture a session", etc. Full reference: references/trace-recording.md.

1. Confirm target — attach to a running app, launch an app, or record all processes? If the user didn't say, ask. List connected devices when useful:

   python3 "${SKILL_DIR}/scripts/record_trace.py" --list-devices

2. Pick a template based on target kind — the SwiftUI template populates the SwiftUI lane on any real device: a physical iOS/iPadOS device or the host Mac. The only exception is the iOS Simulator, where the SwiftUI lane comes back empty — switch to --template "Time Profiler" in that case (still gives Time Profiler + Hangs + Animation Hitches). Always check --list-devices: simulators kind → Time Profiler; devices kind (real devices and the host Mac) → default SwiftUI. Full decision table in references/trace-recording.md.
3. Start the recording. For agent-driven sessions where the user says "I'll tell you when I'm done", start in the background and use a stop-file:

   python3 "${SKILL_DIR}/scripts/record_trace.py" \
       --device "<name|udid>" --attach "<AppName>" \
       --stop-file /tmp/stop-trace --output ~/Desktop/session.trace

   For interactive sessions, just tell the user to press Ctrl+C when done.
4. Signal stop — when the user says they've finished exercising the app, touch /tmp/stop-trace. The script cleanly SIGINTs xctrace and waits up to 60s for finalisation.
5. Analyse the resulting trace (flow into the "Trace-driven improvement" workflow below).

Trace-driven improvement (Instruments .trace provided)

Trigger whenever the user's request references a .trace file. A target SwiftUI source file is optional — if given, cite specific lines; if not, recommend where to look based on view names and symbols the trace already reveals.

Full reference: references/trace-analysis.md. Summary of the composition pattern:

1. Scope the analysis. Ask yourself: does the user want the whole trace, or a slice?
   • "focus on X / after X / between X and Y / during X" → resolve to a window first (see step 2).
   • No scoping cue → analyse the whole trace.
2. Resolve a window (only if the user scoped). The parser exposes two discovery modes:

   # Find a log that marks the start/end of the region of interest:
   python3 "${SKILL_DIR}/scripts/analyze_trace.py" --trace <path> \
       --list-logs --log-message-contains "loaded feed" --log-limit 5
   # Or list os_signpost intervals (paired begin/end), filterable by name:
   python3 "${SKILL_DIR}/scripts/analyze_trace.py" --trace <path> \
       --list-signposts --signpost-name-contains "ImageDecode"

   Both modes accept --window START_MS:END_MS to scope discovery. Pick the time_ms (for logs) or start_ms/end_ms (for signposts) that match the user's description. Build a window like --window 10400:11700.
3. Run the main analysis (with or without --window):

   python3 "${SKILL_DIR}/scripts/analyze_trace.py" --trace <path> \
       --json-only --top 10 [--window START_MS:END_MS]

4. Interpret with references/trace-analysis.md — key diagnostics:
   • main_running_coverage_pct inside each correlation (<25% = blocked; ≥75% = CPU-bound).
   • swiftui-causes.top_sources reveals why updates keep happening — high-edge-count sources like UserDefaultObserver.send() or wide EnvironmentWriter entries are structural invalidation bugs. Fixing one often collapses many downstream hot views.
5. When a specific view shows as expensive, ask who's invalidating it. Use --fanin-for "<view name>" to get the ranked list of source nodes driving the updates.
6. Optionally ground in source. If the user pointed at a file, read it and match view names / user-code symbols against identifiers there. If not, recommend which files to open based on the view names SwiftUI reported.
7. Return a prioritised plan. Cite evidence (coverage %, hot symbol, overlapping view, log timestamp, cause-graph edges) and route each recommendation to a Topic Router reference.
8. Only edit code if the user asked for edits.