Performance profiling skill: measure, find the hotspot, prove the fix

Performance work goes wrong when changes are made on intuition. This skill makes Claude Code follow a measure-change-measure loop and refuse to optimize without data.

**Workflow** 1. **Define the target**: what is slow, how it is triggered, and the current number (latency, throughput, memory, run time). It writes a small repeatable benchmark or reuses an existing one. 2. **Profile** with a tool that fits the stack, for example `node --cpu-prof`, `python -m cProfile`, `py-spy`, Go's `pprof` via `go test -cpuprofile`, or Java Flight Recorder. For browser work it explains how to record a DevTools Performance trace, since it cannot capture one itself. 3. **Read the profile**: lists the top functions by self and cumulative time and connects them to code. 4. **Hypothesize and change one thing**: N+1 queries, repeated parsing, unnecessary allocation, missing index, synchronous I/O, quadratic loops, oversized payloads. 5. **Re-measure** with the same benchmark and inputs, several runs, and report the median.

**Output** A short report: benchmark command, before and after numbers, the profile evidence, the change and why it helped. If the change did not help, it reverts it and says so.

**What it will not do** It does not make micro-optimizations without profile evidence, trade readability for gains within noise, or claim improvements from a single run. It will not run load tests against shared or production environments.