Skill packages
See what Workbench installs, clones, and keeps outside the Agent Skill package.
Workbench publishes and installs standard Agent Skills. A Skill package is what an agent can use directly: SKILL.md plus optional scripts, references, assets, dist/**, and support files.
Workbench adds evals, evidence, versions, improvement loops, and publishing around that package. Those project files are not part of the installed Agent Skill.
Package boundary
| Agent Skill package | Workbench project |
|---|---|
SKILL.md instructions | .workbench/eval.yaml grading standard |
scripts/, references/, assets/, dist/** | .workbench/cases/** workflow cases |
| Files an agent needs at runtime | .workbench/agents.yaml agent configuration |
| Installable published package | Runs, traces, artifacts, results, versions, and lineage |
Keep reusable workflow instructions, scripts, reference material, and assets in the package. Keep evaluation criteria, cases, runtime configuration, and generated evidence in .workbench/**.
Install
workbench skill install copies the Agent Skill package only:
It does not copy .workbench Eval controls, run history, traces, artifacts, remotes, sync state, or local runtime files. For Workbench-published packages, Workbench records the Skill handle and exact published version. External Agent Skill locations can still install through Workbench. Workbench-only behavior such as clone, Eval evidence, improvement lineage, and Cloud visibility does not apply.
Clone
workbench skill clone creates an editable Skill in a fresh Workbench project:
Use install when the recipient only needs the Skill in an agent. Use clone when they need an editable Skill, Evals, and future improvement loops.
Authoring
Use the upstream skill creator best practices for general Agent Skill authoring. Use Workbench when you need a measured development loop: representative cases, graded evidence, version history, improvement proofs, publishing, install, and clone.