The problem: A plan reads fine in the moment you write it. The gap you didn’t think to ask about, the assumption you didn’t know you were making, the domain you didn’t consider: none of that shows up until you’re three milestones in and it’s expensive to unwind. Auditing code after it’s built catches bugs. Auditing a plan before it’s built catches the thing that would have made the whole build go differently.
This is the full gauntlet: ten passes over a plan before it ever becomes code. An interview surfaces what’s in your head that the plan itself doesn’t capture. A principles read checks it against your own standards. A multi-domain audit runs product, UX, security, and ten more lenses against the plan, not the code, so the wisdom gets front-loaded instead of discovered mid-build. A steelman routes the plan through two different external models for a real adversarial pass; Claude synthesizes, it never grades its own homework. A pre-mortem asks what kills this in six months. It converges, round after round, until nothing new survives, capped at three rounds so it can’t loop forever.
Part of Operator Skills, a collection of skills for running a real Claude Code setup.