Lean, bullet-driven prose that assumes Claude's competence — it never explains what plagiarism or a capture pack is, and every section (registry, inventory, comparison, triangulation, fixes, reporting) earns its tokens, matching the level-3 'every token earns its place' anchor.

Provides copy-paste-ready executable guidance — the full `python <skill-dir>/scripts/build_evidence_inventory.py --site ... --reference ... --output ...` invocation plus concrete git commands (`git log -S`, `git log --all --name-status`) and enumerated category checks — rather than vague direction.

A clearly sequenced six-step process with explicit validation checkpoints for a batch/evidence operation — 'Stop and report an access gap when a promised reference is missing', 'Do not clear a site after checking only its homepage screenshot', 'Never turn missing evidence into a pass' — plus a terminal 'Completion checks' checklist, satisfying the level-3 anchor.

SKILL.md is an overview that clearly signals a single one-level-deep reference ('Read [references/audit-rubric.md] before judging findings') and one helper script, both verified to exist; the detailed severity/evidence rubric is correctly split into the reference file rather than inlined, matching the level-3 anchor.

Lists multiple concrete actions — 'Audit a website', 'compare current or historical site output', 'raise evidence-backed red flags', 'distinguish common visual grammar from distinctive copying', and 'propose concrete fixes' — matching the level-3 anchor rather than the single-action level-2 example.

Explicitly answers both what ('Audit ... for originality and plagiarism risk') and when ('Use when Codex must compare ...'), with an explicit 'Use when' trigger clause, so it is not capped at 2 for a missing trigger.

Covers natural terms a user would say — 'reference pages, capture packs, screenshots, copy, brands, numbers, images, assets, videos, layouts, motion, or code' plus 'plagiarism risk' and 'originality' — giving broad coverage of common variations rather than the sparse level-2 'Works with PDF files'.

The niche — auditing a site against its supplied reference corpus for copying — is distinct from generic review skills, with specific triggers like 'capture packs', 'reference pages', and 'originality matrices' making wrong-skill activation unlikely.