Setup

Overview

Find the path to create, evaluate, improve, and publish a Workbench skill.

Workbench mines reviewed workflows from Sources, turns them into Evals, and uses those Evals to improve and publish versioned Agent Skills.

Choose a path

GoalStart here
Ingest evidence and discover workflowsSources
Build your first measured skillQuickstart
Learn the eval source modelEvaluation basics
Author cases and gradingCases and grading
Configure agents and modelsAgents and models
Choose tools for Office, PDF, and other file artifactsFile formats
Improve from failed or below-perfect evidenceImprove from evidence
Review candidate changesReview candidate changes
Rerun proof evidenceRerun proof evals
Inspect resultsResults
Follow runs and jobsRuns and jobs
Inspect traces, artifacts, and filesTraces, artifacts, and files
Review versions and lineageVersions and lineage
Publish a Skill packagePublish
Install or clone a skillInstall and clone
Manage visibility and hosted operationsVisibility and Cloud
Look up command behaviorCLI reference
Understand installable skill packagesSkill packages

Core concepts

ConceptMeaning
SourceAn ingested evidence corpus with immutable snapshots, grounded workflow Analyses, and explicit human review.
EvalThe grading standard for a workflow: cases, grading config, agents, and runtime environment source under .workbench/.
SkillA standards-compliant Agent Skill package with SKILL.md plus support files.

Runs are Eval evidence; package versions and handles are Skill lifecycle records, not additional product roots.

Daily loop

bash
workbench skill new ./earnings-prepcd ./earnings-prepworkbench eval case draft investor-focusworkbench eval runworkbench eval gradeworkbench eval resultsworkbench skill improveworkbench skill publish

Start with Quickstart, then use the pages above for the Source, Eval, or Skill work you need to finish.