Requesting Red Team Review

This is a structured way to catch mistakes before you commit to a conclusion. You dispatch a separate subagent with just your analysis artifacts (bounded by git SHAs, not your full session history) and an explicit mandate to find what's wrong. The reviewer looks for confounds, leakage, assumption violations, and over-claiming. It's meant to be mandatory after completing analysis steps and before reporting findings, especially when results are suspiciously clean or convenient. The pattern borrows from adversarial review in research: a conclusion that survives a genuine attack is worth trusting, one that hasn't been attacked isn't. Critical issues block reporting, Important issues get fixed before writeup, and you're expected to verify rather than dismiss feedback that stings.

Install to Claude Code

npx -y skills add k-dense-ai/science-superpowers --skill requesting-red-team-review --agent claude-code