Use when the user asks to audit, falsify, red-team, pressure-test, or verify a theorem, chapter, proof, compute module, or frontier claim in this repository. Not for straightforward local edits that do not require an adversarial audit.
Run this skill when correctness matters more than speed and the task is to challenge a claim, not merely improve its presentation.
CLAUDE.mdchapters/connections/concordance.texmetadata/theorem_registry.mdarchive/raeeznotes/raeeznotes100/red_team_summary.mdFor status-heavy work, also load the relevant metadata files.
CRITICAL: false theorem surface, circular proof, or proved-here claim leaning on weaker materialSERIOUS: wrong formula, wrong scope, wrong object, unstable dependency, or convention errorMODERATE: misleading prose, missing qualification, stale status sync, or incomplete verificationMINOR: wording, exposition, or low-risk cleanupIf the user asked for a review or audit, present findings first, ordered by severity, with exact file references.
If no findings survive, say so explicitly and still mention residual risks or verification gaps.
Only use subagents if the user explicitly asks for parallel or delegated agent work. If that happens, split audits by independent scope, not by overlapping edits.