---
name: benchmarking
description: A set of guidelines to use when asked to benchmark
compatibility: opencode
---

Use this skill when the user asks you to benchmark some code, infrastructure or an approach.

It's important you do not just run the test once. At least 5 tests should be run, ideally 10. Then median and p95 results should be obtained to give the user a full picture of the implications.

If applicable, do separate tests for cold/pre-warmed outcomes. It's important that you give the user a full picture of the performance, especially if doing before/after testing in order for them to have a full picture.