Conciseness

The body is dense and reference-shaped — tables carry the payload with almost no prose padding and no explanation of concepts Claude already knows — but there is real duplication: the "Deep references" and "References" sections list the same four reference links twice, and "How to use this reference" re-enumerates pattern names ("cache-aside, write-through, write-back, write-around, refresh-ahead") that the tables immediately below define again. That places it at the score-4 anchor ('efficient; minor instances that could be trimmed'), above score-3 ('some unnecessary explanation') but short of score-5 ('every token earns its place').

Actionability

For an instruction-only reference skill the guidance is fully concrete: exact directives ("Cache-Control: private", "Vary: Authorization", "s-maxage longer than session lifetime"), exact fixes in the anti-patterns table ("no-store, no-cache, must-revalidate, private"), and a worked example that ends in a copy-paste-ready test shape with Arrange/Act/Assert steps and concrete failure diagnoses ("A failure here means the write path skipped the invalidate"). Per the code-vs-instruction scoring note, absence of code is not penalized when guidance is this actionable, matching the score-5 anchor ('specific examples cover the common cases').

Workflow Clarity

"How to use this reference" gives a clear 5-step sequence (locate tiers → choose write pattern → choose invalidation strategy → set the contract → write the per-tier coherence test) that ends in an explicit validation checkpoint: "Write the per-tier coherence test... then re-check the design against the Anti-patterns table below." The anti-patterns table functions as a checklist and the worked example includes error-recovery guidance, matching the score-5 anchor ('explicit validation steps; feedback loops; checklists') rather than score-4 ('most checkpoints present; minor validation gaps').

Progressive Disclosure

The body behaves exactly as an overview should: it keeps the decision surface (tier table, pattern tables, anti-patterns) inline and defers deep detail to five real, one-level-deep files, each surfaced with a descriptive bullet ("RFC 9111 directive tables", "Cache stampede (thundering herd) - symptoms, the three mitigation families, the XFetch formula"). All referenced paths exist in references/, and the reference files' lateral cross-links never bury content two hops from SKILL.md. This matches the score-5 anchor ('clear overview with well-signaled one-level-deep references; content appropriately split').

Specificity

The description lists multiple concrete capabilities and contents: "Defines the five-tier cache stack (browser → CDN → reverse-proxy → application → data store)", enumerates all five cache-writing patterns and five invalidation strategies by name, and details what lives in references/ ("RFC 9111 Cache-Control / Vary / ETag directive tables", "the XFetch formula", "RFC 5861 stale-while-revalidate / stale-if-error semantics"). Coverage is comprehensive with no generic filler, matching the score-5 anchor rather than the score-4 anchor ('several specific actions; minor gaps').

Completeness

It explicitly answers both questions: what — "Pure-reference catalog of cache-coherence patterns... Defines the five-tier cache stack... the per-tier cache-writing patterns... the canonical invalidation strategies" — and when — "Use for pattern selection, Cache-Control header design, coherence audits, stampede-refresh strategy, and SWR/SIE window design". This mirrors the score-5 anchor's structure (explicit what + explicit when with concrete trigger phrases), not the score-4 anchor where the 'when' could be more specific.

Trigger Term Quality

Good keyword coverage including synonyms and abbreviations users actually say: "cache-stampede (thundering-herd)", "stale-while-revalidate / stale-if-error", "SWR/SIE", "Cache-Control header design", "event-driven purge", "surrogate keys". A few natural phrasings are missing — e.g. users reporting "stale data", "cache invalidation", or "CDN caching" would not literally match — so it sits between the score-4 and score-5 anchors, clearly above the score-3 anchor ('missing common variations or synonyms').

Distinctiveness Conflict Risk

It carves a clear niche (cross-tier cache coherence patterns) and even disambiguates against a sibling skill: "use a cache-key-collision check when the question is whether two concrete requests collide on a key scheme". Trigger terms (coherence audits, stampede mitigation, SWR/SIE windows) are unlikely to fire for unrelated skills, matching the score-5 'clear niche with distinct triggers' anchor rather than the score-4 'minor overlap risk with closely related skills' anchor.