kayforkind/skill-slice/android-performance
Claude Code and agent skill by kayforkind/skill-slice
- community
Checking that the SKILL.md file is still in the repository.
Captures Android performance evidence on an adb target: Simpleperf CPU samples, Perfetto/Compose traces, gfxinfo framestats, dumpsys meminfo, Java heap dumps, and native heapprofd allocations. Use when profiling a focused Android app flow, diagnosing jank or startup stalls, ranking CPU hotspots, or gathering leak artifacts. Not for unfocused soak captures, installing/driving the UI, or treating empty Simpleperf output on a non-profileable build as proof.
Install
npx vskill@latest install kayforkind/skill-slice/android-performanceSKILL.md
The same SKILL.md was published earlier as kayforkind/skill-slice/android-performance.
Android Performance
Capture and interpret Android performance evidence for adb-installable apps. CPU sampling usually requires a debuggable or profileable build; frame stats, Perfetto, and logcat can still help when an app cannot be sampled. Compose with ../android-emulator-qa/SKILL.md for device selection, build/install/launch, UI driving, screenshots, UI trees, and logcat capture.
When to Use
Use this skill when any of the following come up:
- "Profile this Android app flow" or "what functions are taking CPU time?"
- Diagnosing jank, missed frames, slow startup, or main-thread stalls
- Capturing frame timing or scheduler/binder/lock evidence
- Comparing before/after performance of a code change
- Gathering memory snapshots, Java heap dumps, or native allocation traces
- Explaining what code is taking time during a user-visible flow
Do not use for broad "use the app for a while" captures — they make traces hard to attribute and hide the functions that matter. Always pick one focused user-visible flow.
Prerequisites
- A local adb target (device or emulator) with the app installed and reachable via
adb -s <serial>. - For Simpleperf CPU sampling: the installed package must be debuggable or profileable from shell. Frame stats, Perfetto, and logcat work regardless.
ARTIFACT_DIR— a run-specific output folder outside the skill directory. The skill folder is for bundled instructions and scripts, not run artifacts.- Optional: Shark CLI for Java heap dump analysis;
trace_processor_shellfor Perfetto native allocation analysis. - Windows host is primary (PowerShell). When running bash snippets on Windows, use Git Bash, WSL, or an equivalent bash shell. Keep Windows path notes in mind when constructing
SKILL_DIRpaths.
Procedure
0. Set up artifact directory
HARD RULE: Do not putARTIFACT_DIRunderSKILL_DIR. The skill folder is for bundled instructions and scripts only.
1. Choose a trace type
| Question | Trace type | |
|---|---|---|
| What functions are taking CPU time? | Simpleperf | |
| Frame timing, startup timeline, scheduler gaps, binder, locks, main-thread stalls, Compose recomposition, jank | Perfetto | |
| Quick manual frame/jank snapshot | gfxinfo framestats (pair with Perfetto for root cause) | |
| Retained Java/Kotlin objects, PSS, native heap, object counts | meminfo / heap dumps |
2. Simpleperf CPU Profiles
2a. Preflight — verify debuggable/profileable
If the package is not debuggable/profileable and simpleperf record --app fails, install a debug/profileable build when possible. If that is not possible, use Perfetto or gfxinfo instead of treating missing CPU samples as evidence.
2b. Start recording
While that command is running, perform exactly one focused flow with adb input, UI automation, or android-emulator-qa.
2c. Stop recording
If that returns Operation not permitted, send Ctrl-C to the original adb shell simpleperf record command session and wait for it to exit.
2d. Pull and report
HARD RULE: Do not deriveSKILL_DIRfrom the target app repo'spwd; installed plugins usually live outside the app being profiled.
HARD RULE: KeepFIRST_PARTY_REGEXscoped to the app's package or app-owned module prefixes. Avoid broad framework patterns such askotlin,Compose, orandroidx.composewhen reporting app-owned rows.
The helper writes:
$ARTIFACT_DIR/simpleperf-self.txt$ARTIFACT_DIR/simpleperf-children.txt$ARTIFACT_DIR/simpleperf.csvwhen supported by the installed Simpleperf
If host Simpleperf is not installed, the helper searches Android Studio and Android SDK/NDK locations. If unavailable, it falls back to device-side adb shell simpleperf report when the device still has /data/local/tmp/perf.data.
2e. Reading Simpleperf reports
Simpleperf reports sampled CPU execution. It does not directly measure suspended coroutines, network latency, lock wait time, or other wall-clock waits. If a flow feels slow but Simpleperf shows little app CPU, capture Perfetto.
- Self/Overhead: samples where the function itself was executing. Use for hot leaf work — parsing, formatting, diffing, sorting, allocation-heavy iteration, JSON/protobuf processing.
- Children/inclusive: samples in the function and its callees. Use for expensive entry points — repositories, use cases, view models, Composables, startup initializers, feature coordinators.
- Shared Object / Symbol: prefer app-owned package frames, feature modules, domain/data/UI modules, and generated app code. Treat Android framework, Kotlin runtime, Compose, and native/runtime frames as context unless the app-owned caller is visible.
- Percentages: useful for ranking functions inside one capture. For user-facing timing claims, pair with Perfetto,
gfxinfo, or repeated wall-clock measurements.
When interpreting a hotspot, note: symbol/function name, self or inclusive percentage, approximate sampled CPU time when available, caller stack or owning source file, flow steps, artifact paths, and whether the capture is single-run or repeated.
3. Perfetto / Compose Trace
If the app repo already documents a Perfetto/System Trace command for that project, use it. Otherwise use the light command below. It captures scheduler/frequency/Android atrace categories and app Trace sections for PACKAGE; it is not a substitute for a full project-specific Perfetto config when you need detailed frame timeline or Compose runtime internals.
3a. Start Perfetto
Run exactly one focused flow before TRACE_DURATION_SECONDS expires.
3b. Stop early (optional)
Prefer letting TRACE_DURATION_SECONDS expire instead of stopping early. To stop early:
If the stop command fails because the trace already ended, still wait until the output file exists and its size is stable before pulling.
3c. Pull the trace
3d. Inspect in Perfetto
Load references/ for Perfetto query examples when available. In Perfetto UI or trace_processor_shell, inspect:
- Main-thread slices around missed frames or long startup sections
- Frame scheduling, frame timeline, and render thread lanes
- Compose runtime tracing sections for recomposition work when enabled
- Binder transactions, monitor contention, scheduler gaps, and app log markers
Only report frame timeline or Compose recomposition details when those tracks/events are actually present in the captured trace. The light command above does not guarantee them.
4. gfxinfo Framestats
Use for a quick manual frame/jank snapshot:
Capture from a stable, responsive screen. If dumpsys gfxinfo fails to dump the process, or the device shows an ANR/dialog/splash screen instead of the flow, discard that capture and use Perfetto for root cause.
Read the headline summary first: total frames, janky frames, frame percentiles, slow UI thread, slow draw commands, and frame deadline misses. On emulators, absolute smoothness numbers are noisy; percentile spikes and slow draw/UI counters are still useful for deciding whether to take a Perfetto trace.
5. Memory / Leak Artifacts
Use on an adb target after narrowing the investigation to one flow. Exercise the flow, return to a stable screen, then capture memory artifacts from that state.
5a. meminfo snapshot
Read TOTAL PSS, Java heap, native heap, graphics, Views, Activities, binder counts, and object counts. Treat one noisy sample as a lead, not a conclusion.
5b. Java heap dump (Shark CLI)
For retained Kotlin/Java objects, prefer Shark CLI when available. It works with Android heap dumps and produces text output the agent can inspect and cite.
Read shark-analysis.txt first when it exists. Report suspected leaking objects, retained sizes, and reference chains. Look for retained feature objects, activities, fragments, view models, Compose state holders, repositories, listeners, callbacks, and caches that should have been released after leaving the flow. If Shark CLI is unavailable, still preserve the HPROF path and inspect it with the best available heap analyzer; do not claim leak roots from meminfo alone.
5c. Native allocation traces (heapprofd)
For native allocation growth, capture a Perfetto trace with heapprofd enabled.
HARD RULE: Current Androidperfettorejects-ttogether with--config. Keep the duration in the config.
This is the start of the file. Read the full SKILL.md on GitHub.
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/android-performance/SKILL.md
Popularity
- 7-day trend score
- 5
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/kayforkind/skill-slice/android-performance)