loop-until-done

The canonical loop skill. Two modes, one entry point:

Mode 1: Drive-to-completion (default)

PRD-driven execution — breaks work into stories with acceptance criteria, iterates story-by-story with independent verification, terminates only when every criterion has fresh passing evidence.

Phases:

1. PRD planner — generates stories with testable acceptance criteria
2. Falsifiability judge — marks each criterion pass/fail for falsifiability
3. Executor — completes each story, runs verification
4. Verifier — per-criterion evidence check
5. Reviewer — final verdict with configurable reviewer (--critic)

From ralph: session persistence, structured story tracking (prd.json), progress tracking across iterations, reviewer verification before completion.

From autopilot: autopilot should invoke this skill for its Phase 3 verify loop rather than implementing its own.