from crucible10
Adversarial review skill: iteratively attack artifacts (designs, plans, code, docs) and surface fatal/significant issues until clean or escalation.
The Red Team skill performs adversarial review of artifacts (design docs, implementation plans, code, pull requests, documentation). It dispatches a Devil's Advocate subagent to identify Fatal, Significant, and Minor issues, computes weighted scores (Fatal=3, Significant=1), and can iterate with fix agents until the artifact is clean or stagnation is detected. The skill enforces fresh reviewers each round to avoid anchoring.
Invoke this skill when you need a rigorous adversarial analysis before major milestones: before planning, before finishing an implementation, or as part of a quality gate. It’s suited for safety-critical paths, launch checks, and anywhere hidden assumptions or failure modes matter. For CI-driven quality gates, use single-pass mode; for standalone invocation, the skill can coordinate the full iterative loop.
Best used with agent runtimes that support subagent dispatch and orchestration (e.g., Agent Zero, Claude Code workflows, Cursor).
Red-team skill for adversarial review of artifacts (designs, plans, code, docs). Dispatches Devil's Advocate subagents iteratively with stagnation detection via weighted scoring. No bundled scripts — pure instruction-based skill. Well-structured with clear triggers, dual-mode operation, and anti-rationalization safeguards. Niche but valuable for teams using the Crucible framework.
No security concerns — no scripts, no network calls, no credentials. Pure prompt/instruction skill. Well-organized frontmatter and clear structure. The skill is tightly coupled to the Crucible ecosystem and won't work without it, limiting broad usefulness. Depth calibration table and rationalization prevention are thoughtful touches.
