What a team is Shipped
A directory of markdown and JSON. No Python, no engine hooks. That constraint is what makes a team something you can author, judge, improve and delete.
A team is data. It declares who its specialists are, what each one is expert in, what commands it answers to, and which engine workflow each command binds to. It ships no code at all — if a design needs behaviour the engine does not have, that is a gap to close in the engine, not something a team papers over.
What is inside one
- team.jsonIdentity, commands, and which engine each command binds to.
- team.mdThe team document — its role, and the specialists that make it up. Some teams add their calibration and roster rules here too.
- specialists/One directory per persona, each declaring the
specialities:it works through. - specialities/The lens bodies — Worker Focus, and the criteria a verifier checks.
- evals/The rubric it is judged against, shipped with the team it judges — where one exists.
Only team.json and team.md are obligatory. teamtester is the proof: it declares no specialists and no specialities at all, because it tests other teams rather than reviewing anything itself. And only two of the nine — devteam and socialteam — currently ship an evals/ rubric, so the improvement loop below has a target on two teams and nothing to score the other seven against yet.
Why "ships no Python" is the load-bearing rule
It is what makes a team droppable. Adding one is a directory plus a registry entry; deleting one touches no engine code and breaks no other team. It is also what makes teambuilder possible at all — generating a team is generating markdown, which is a thing a model can be held to a template on.
Four things every team gets for free
- A terminal front door —
myteams <team> <command>. - A Claude Code skill —
/myagenticteams:<team> <command>, regenerated on every registry change. - A run record — an event store, so the run can be replayed, reported on, and probed afterwards.
- The six shared run commands, if it declares a
report— report, pause, resume, log, findings, insights.
How many teams produce runs
brainstormteam is the interactive one: it produces a journey rather than a run with findings, so it declares no report and inherits none of the six. That is the composition rule working, not an omission.