The body assumes Claude's competence and avoids explaining basics; every section (CLAIM/EXTRACT/DOUBT/RECONCILE/STOP) earns its tokens. Minor repetition in the cross-model section keeps it just at the lean anchor rather than above it.

Provides a verbatim adversarial prompt, a copy-paste checklist, concrete bash invocations with stdin piping to avoid shell escaping, and a precedence-ordered classification scheme — fully executable guidance.

The five-step process is clearly sequenced with explicit validation checkpoints (re-read artifact before classifying), a bounded loop with stop conditions, escalation after 3 cycles, and a final verification checklist.

The body references references/orchestration-patterns.md and agents/ as if they exist, but neither bundle path is present in the skill directory — these broken references prevent the one-level-deep navigation the top anchor requires.

"Subjects every non-trivial decision to a fresh-context adversarial review before it stands" names a single concrete action with a specific mechanism, matching the 'lists multiple specific concrete actions' anchor.

It explicitly states both what it does (subjects decisions to adversarial review) and when to use it via an explicit "Use when..." clause, satisfying both halves.

"Use when correctness matters more than speed, when working in unfamiliar code, when stakes are high (production, security-sensitive logic, irreversible operations)" gives broad natural-language trigger coverage a user would actually say.

The "fresh-context adversarial review" niche and explicit contrast with /review make it clearly distinguishable and unlikely to trigger for the wrong skill.