Red Team анализ планов и архитектуры

★ 7.7 · research

adversarial is a Claude Code skill that applies opposite-analysis to plans, specs, architecture, code changes, and claims — surfacing hidden assumptions, failure modes, and risks through structured red-team reasoning. The core six-step workflow states the challenged claim, maps its assumptions, tests what would make each assumption false in the current repo or runtime, ranks failure modes by impact, separates proven issues from plausible risks, and ends with a single concrete recommendation: keep, adjust, defer, validate, or revert. For non-trivial work, it runs an Aristotle First Principles pass across five phases and spawns parallel subagents — security, reviewer, and tester — each forming its own position without seeing the others' reasoning. Security mode covers auth, input validation, network boundaries, data exposure, and supply chain risks, grounding every finding in file paths, code behavior, or runtime evidence. The main agent retains final ownership of decisions and synthesis.