ralph

A PRD-driven persistence loop verifies acceptance criteria by repeatedly implementing code until every user story passes. It tracks story priority, gathers fresh evidence for each criterion, and requires reviewer sign-off before completion to eliminate done-but-wrong outcomes.

Preventing done-but-wrong outcomes requires mandatory reviewer approval passes using configurable tiers like architect, critic, or codex. A PRD-driven loop re-validates each story's acceptance criteria with fresh evidence before allowing completion sign-off.

Automating PRD scaffolding and refinement is possible by maintaining a structured prd.json of user stories and acceptance criteria. The system auto-scaffolds the PRD when missing and can add new stories dynamically when sub-tasks are discovered during execution.

Running cleanup and regression checks involves executing an ai-slop-cleaner pass on changed files followed by relevant test or build checks. This ensures the cleanup process does not introduce regressions and maintains passing acceptance criteria.

Multi-agent orchestration handles story-by-story execution by selecting the next highest-priority user story with a false pass status, updating progress tracking, and iterating until all stories pass verification with reviewer sign-off.

A PRD-driven verification loop is not ideal for simple, single-pass tasks that lack defined acceptance criteria or user stories. It is designed for multi-iteration engineering tasks where partial failures require persistence and re-verification.