PAUL (Plan-Apply-Unify Loop)
PAUL forces a Plan-Apply-Unify loop onto Claude Code: you define the work, the agent runs it task by task with a status at each step (done, done with concerns, needs context, blocked), then it reconciles what was planned against what actually happened. Distributed on npm (`paul-framework`), installable via `npx paul-framework`.
Strengths
- The Unify step reconciles planned vs actual and catches false completions before they pile up
- Explicit per-task statuses (done / done with concerns / needs context / blocked) instead of a binary done/not-done
- npm-distributed, installable without cloning: `npx paul-framework` and the loop is set up
Limitations
- Largely overlaps spec-kit and RPI methods: stacked on an existing harness, it duplicates the Plan-Apply part
- Overkill for small tasks: the full loop only pays off on work big enough to risk false completion
Best for
- Dev or tech lead steering a large multi-task job wanting per-criterion post-execution verification
- Team tired of agents declaring 'it's done' on incomplete work and looking for a guardrail
- Builder wanting to compare spec-before-code disciplines before choosing their own