skills

Write Java microbenchmarks with JMH (Java Microbenchmark Harness) that produce trustworthy numbers — not numbers distorted by JIT dead-code elimination, constant folding, insufficient warmup, or single-fork JIT contamination. Use this skill whenever the user writes `@Benchmark`, mentions JMH, microbenchmark, throughput measurement, latency measurement, or compares two implementations performance-wise. Triggers on `@Benchmark`, `@Setup`, `@State`, `Blackhole`, `@Fork`, `@Warmup`, `@Measurement`, `@OperationsPerInvocation`, `Mode.Throughput`, `Mode.AverageTime`, `Mode.SingleShotTime`, `Mode.SampleTime`, `BenchmarkMode`, `OutputTimeUnit`, `BenchmarkRunner`, `jmh-core`, `jmh-generator-annprocess`, `org.openjdk.jmh`, `me.champeau.jmh` (Gradle plugin), `pl.allegro.tech.build.axion-release`, `JMHTask`. Treat any benchmark without `Blackhole.consume()`, with `@Fork(0)`, or with `@Warmup(iterations < 5)` as broken by default — pre-flight statically before running. Pair with `async-profiler` (`-prof async`) for per-ben

Profile JVM applications with async-profiler — sampling-based CPU, allocation, lock, wall-clock, and hardware-counter profiling that produces flame graphs, JFR, and pprof. Use this skill whenever the user investigates JVM performance, mentions hot methods or hot paths, asks about flame graphs, JFR, async-profiler, asprof, libasyncProfiler, jfrconv, or AsyncGetCallTrace, debugs allocation pressure or GC churn, hunts lock or monitor contention, traces latency spikes or time-to-safepoint issues, profiles a Java/Kotlin/Scala/Clojure process, sets up production-safe continuous profiling, or compares before/after performance with differential flame graphs. Use it even when the user only describes the symptom (e.g., "my JVM uses 100% CPU", "young GC is firing constantly", "threads are stuck") without naming the tool — async-profiler is the default JVM profiling choice over JFR-only and VisualVM.