Closing the loop on AI-assisted development drift

What it does

PAUL is a command framework that layers project-management discipline over Claude Code. It prompts you to define acceptance criteria in Given/When/Then format before any work starts, then tracks execution through a mandatory three-phase loop: Plan, Apply, Unify. State persists across sessions in a local .paul/ directory, so resuming work means reading STATE.md rather than reconstructing context from scratch.

The interesting bit

The framework deliberately resists the usual subagent sprawl: implementation stays in-session where context is richest, while subagents are relegated to research and discovery. A forced UNIFY phase reconciles what was planned against what actually happened, producing a SUMMARY.md and updated state—no orphan plans allowed. It is essentially an opinionated workflow encoded as slash commands and markdown artifacts.

Key highlights

• Scope-adaptive planning that auto-routes work as quick-fix, standard, or complex ceremonies
• Execute/Qualify loop with escalation statuses like DONE_WITH_CONCERNS and BLOCKED instead of naive pass/fail
• Anti-rationalization enforcement to prevent false completion claims
• Optional BASE v2 integration turns PAUL projects into queryable nodes in a workspace knowledge graph
• Works for non-code projects—marketing campaigns, workflows, and automations are first-class citizens

Verdict

Builders who treat Claude Code as a daily shipping partner and are tired of debugging AI output should look here. If you prefer ad-hoc prompting without acceptance criteria or session bookkeeping, this will feel like bureaucratic overhead.