Skip to content

Agent skill

BenchBro ships an agent skill with the Python package. It gives compatible coding agents version-matched guidance for defining benchmarks, choosing measurement settings, managing baselines, interpreting results, and using advanced features.

Install it in a project

Install BenchBro first, then run Library Skills from the project root:

$ uv add --dev benchbro
$ uvx library-skills --skill use-benchbro --yes

This discovers the skill embedded in the installed BenchBro distribution and links it into the project's .agents/skills/ directory. Because the link points at the installed package, upgrading BenchBro also updates the skill content.

Run the interactive command when you want to choose from every skill exposed by a project's dependencies:

$ uvx library-skills

Use uvx library-skills --check to verify managed links without changing them.

Invoke it

Compatible agents can select the skill automatically when a task imports or uses BenchBro. Invoke it explicitly when you want to ensure the guidance is loaded:

$use-benchbro add a parameterized benchmark for this parser and create a local baseline

The skill treats the installed BenchBro API and CLI as the final source of truth, so it remains useful in editable installations and when working across versions.

What is included

  • the common Case, input, system, parameterization, discovery, and run workflow;
  • reliable-measurement and confidence-aware comparison guidance;
  • local, named, and strict CI baseline behavior;
  • subprocess, isolation, timeout, CPU-affinity, and profiling patterns;
  • pytest and programmatic API integration guidance;
  • common failure modes that can invalidate a benchmark or comparison.

The source lives in .library-skills/use-benchbro/. Package builds map it to benchbro/.agents/skills/use-benchbro/ inside the wheel, which is the location Library Skills discovers after installation.