Workbench
Log inView Workbench on GitHub
Setup

Quickstart

Create a skill, add one case, run an eval, inspect evidence, and publish.

Build a measured skill. Create a project, add one eval case, run it, review the evidence, improve, and publish.

What you'll build

This guide creates:

  • a local Workbench skill project
  • one representative eval case
  • one recorded eval run
  • evidence you can inspect
  • a published handle you can install or clone

Prerequisites

  • Node.js and npm
  • Provider login for provider-backed evals, such as Codex or Claude
  • Workbench Cloud login when you publish or run hosted operations

Install

bash
npm install -g @workbench-ai/workbenchworkbench --help

Provider-backed evals also need the provider CLI login for the agent you use.

Create a skill

bash
workbench skill new ./earnings-prepcd ./earnings-prep

Edit SKILL.md so it describes the workflow the agent should perform. Replace the generated trigger description, inputs, workflow, output, and quality criteria before treating eval results as meaningful evidence.

Add one case

Create a representative case:

bash
workbench eval case draft investor-focus

Edit .workbench/cases/investor-focus/case.yaml. A provider-backed case can be as small as a prompt plus grading criteria:

yaml
prompt: Create an earnings prep note for GOOGL.grade:  with:    criteria:      - id: investor-focus        description: Explains the likely investor focus areas.      - id: supported-facts        description: Avoids claims not supported by the case context.

Run and review

bash
workbench eval runworkbench eval gradeworkbench eval listworkbench eval resultsworkbench eval show RUN_ID

workbench eval run records output before grading. workbench eval grade judges existing output. Workbench keeps the two actions explicit so you can iterate on prompts and judgment independently. workbench eval list lists eval identities, including the current draft and immutable historical eval versions created from authored eval source.

Improve

After below-perfect, failed, or reviewed evidence exists:

bash
workbench skill improveworkbench eval runworkbench eval grade

improve edits the skill package, proves the candidate with eval evidence, and switches only when the proof beats the current version.

Publish and share

bash
workbench loginworkbench skill publish --public

The publish output prints the install command to share:

bash
workbench skill install OWNER/SKILL

Use workbench skill clone OWNER/SKILL ./local-copy when the recipient needs editable source, evals, and future improvements.

Next steps