Break an epic into small, independently-deliverable stories — vertical slices, not a task breakdown. The second decomposition-loop beat: product and architecture questions that only become visible while splitting get surfaced and recorded.
The declared "Claiming the work is done" checkpoint, made real. An epic is not done because its stories are closed — it is done when the PRD's success metrics are each satisfied or explicitly flagged out-of-scope-for-v0, and a human agrees. No agent declares victory alone.
Have the Project Manager estimate the decomposed epic and test it against the stated timeline — pushing back honestly instead of absorbing pressure. The judgment beat: "do we accept this estimate, or change scope/date?" is a human decision.
Walk a PRD with the Product Owner. Surface silent assumptions, internal contradictions, vague success criteria, and the open questions that belong to humans — without answering any of them. The first read-only beat of the workshop.
Turn a reviewed PRD into a technical implementation epic — honouring the architecture constraints in CLAUDE.md absolutely, and stopping at every decision the room owns. The first generative beat, and the first of the two HITL loops (decomposition).
Close a single story honestly: its acceptance criteria are met with evidence, it has passed review, and "done" is a human decision — not an agent's self-declaration. The per-story half of the orphaned "claiming the work is done" checkpoint.
Kick off work on one or more ready stories. This is the second HITL loop — implementation-time. Opening a story routinely exposes architecture or business questions invisible at decomposition. This command surfaces them, records the decisions, and stops at the gate — it does not write production code.
★not rated 5 3mo agoA0 tokens
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: