The adversarial plan review, generalised and ready to open-source
outcome the lead candidate for public, inspectable proof
Problem
The cross-model plan review proven at work was bound to one environment. A transferable version is worth more: to other teams, and as public proof.
What I built
Three self-authored, project-agnostic skills implementing the full pattern: a batched stakeholder interview with a recommended answer per question, the locked plan handed to OpenAI Codex in a read-only sandbox, an explicit APPROVED or REVISE verdict protocol with the same reviewer session preserved across bounded revision rounds, and human sign-off before code.
A docs-aware variant additionally maintains a domain glossary and architecture decision records as decisions crystallise. Built on Matt Pocock’s MIT-licensed grill-me.
Outcome
The pattern is battle-tested on personal projects and is the lead candidate for the planned open-source release: a public, inspectable artifact no blog post can match.