The body is ~25 lines with no padding and no explanation of concepts Claude already knows; every line earns its place and it leans on a reference file for depth, matching the lean/efficient anchor.

Guidance like "Clean up unused tags, use asynchronous loading, and consider moving to a server-side implementation" is concrete in intent but provides no executable code or commands in the body itself (those live in the reference), so it stops at "some concrete guidance but incomplete."

The Check/Fix/Explain/Code Review sections give a loose sequence, but there is no explicit validate-then-fix feedback loop or checkpoint in the body for an auditing task that recommends changes, which caps workflow clarity at 2.

The body is a concise overview with a clearly signaled one-level-deep pointer to references/rule.md (a real file), keeping detail appropriately separated and easy to navigate.

Names multiple concrete actions — "auditing slow page loads, heavy assets, or rendering delays" and "Verify the actual bottleneck in DevTools, Lighthouse, or field data before recommending changes" — matching the anchor listing several specific concrete actions.

Explicitly answers both what (audit and verify GTM performance bottleneck) and when ("Use when auditing slow page loads, heavy assets, or rendering delays related to ... Google Tag Manager"), with an explicit trigger clause.

Uses natural terms a user would actually say — "slow page loads", "heavy assets", "rendering delays", "Google Tag Manager", "DevTools", "Lighthouse" — giving good coverage of common phrasings.

The explicit GTM qualifier carves a clear niche unlikely to trigger for unrelated skills; it sits slightly above level 2 because the performance-symptom triggers are anchored to GTM rather than left generic.