# Performance Investigation

> Investigate latency, throughput, memory, CPU, I/O, bundle size, or scalability using measurement, profiling, experiments, and benchmarks.

- Skill: `egeetas/performance-investigation` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add egeetas/performance-investigation`
- Raw SKILL.md: https://api.skillmd.com/api/skills/egeetas/performance-investigation/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: egeetas (https://skillmd.com/u/egeetas)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/egeetas/performance-investigation

---


# Performance Investigation

Measure before optimizing. A performance suspicion is not evidence of a bottleneck.

## Workflow

1. Define the affected workload, user impact, target metric, percentile, environment, and acceptable threshold.
2. Establish a repeatable baseline with representative data and controlled conditions.
3. Profile the relevant resource: wall time, CPU, allocations, memory growth, queries, network, disk, rendering, or bundle composition.
4. Rank bottlenecks by measured contribution and form falsifiable hypotheses.
5. Change one factor at a time and compare against the same baseline.
6. Check correctness, cold/warm behavior, tail latency, variance, resource trade-offs, and scale characteristics.
7. Add a stable regression benchmark or budget when the project can maintain it.

Report environment, commands, dataset, repetitions, before/after results, uncertainty, trade-offs, and remaining bottlenecks. Do not generalize microbenchmark results beyond their measured workload.