skills-studio

Authoring, evaluation, promotion, and estate scanning. Deliberately the last thing built.

An evaluation platform with nothing to evaluate is a well-known way for a platform team to spend a quarter and ship nothing. The gate cannot be designed correctly until real submissions have shown what actually goes wrong, so the first version of every check is a person doing it by hand and writing down what they looked for.

The four planes

Author wraps {{ORG}}/plugin-template with scaffolding and linting rather than reimplementing it.

Eval runs held-out splits, measures trigger rate across repeated runs, and compares blind against the prior version.

Govern opens the pull request that adds a catalog entry to {{ORG}}/agent-marketplace. The studio never writes to the marketplace directly, so the catalog’s git history stays the audit trail.

Observe tracks installed versions per harness, because Claude Code and Copilot have asymmetric auto-update policies and the two populations drift.

The estate scan

The one capability nothing else provides, and its framing matters more than its implementation.

Skills resolve first-found-wins with plugin directories near the bottom, so a local file shadows a published skill of the same name with no warning. Under a compliance framing that is a leak to close. Under an innersource framing it is a fork, and forks are how the community tells you what the standard got wrong.

Route good forks back as contribution candidates. Default assumption when a collision appears: the local skill is right and the published one needs work.

Deployment

On-prem, single writer. Keep it reconstructible from the marketplace repository plus an evaluation artifact volume, so losing the studio database costs history rather than the platform.

The explorer app

The interactive repository explorer that once lived here as app/ is generated by the companion agent-platform-explorer skill (scripts/explorer app <dir>), which also produces the standalone HTML and A2UI renderings from the same topology.