The problem: A plan you just wrote sounds right on a re-read because you already believe it, and that’s confirmation, not review. The real failure modes, an assumption nobody said out loud, an edge case that only shows up in week three, a defensive feature that’s really just fear wearing a to-do item, all pass a self-read clean, because a self-read never disagrees with you.
This forces a real argument instead: write the plan, critique it like an outsider would, then defend it the way you’d defend a colleague’s work you respected, and only fold in a criticism once it’s survived that defense. In this system’s own use, something close to half of the “critical issues” a first self-critique turns up don’t survive the steelman: fear-driven, already handled elsewhere, or defending against a scenario that structurally can’t happen. The rest are real precisely because they survived you trying to kill them.
Part of Operator Skills, a collection of skills for running a real Claude Code setup.