Conciseness

The body is dense and operational — exact commands, exact issue-line formats, a fix decision table, and templates with no padding or explanation of known concepts. It falls short of anchor 5 because the intro paragraph restates the frontmatter description and the 'When to invoke' section duplicates that trigger guidance again; the install-path parenthetical is also longer than needed. Not anchor 3 — there is no unnecessary explanation to speak of.

Actionability

Everything is copy-paste ready: 'deno run --allow-read scripts/audit.ts <memex-root>/pages/', 'grep -rln "\[!WARNING\]" <memex-root>/pages/', exact per-issue output formats, an exact stub-page template, and full report and log.md templates. The bundled script exists and behaves as documented. Matches anchor 5.

Workflow Clarity

Six clearly numbered steps in a coherent sequence with severity classification, a constraints checklist, and safe default behavior. However, the `--fix` path is a batch write operation (stub creation, index.md edits, section appends) with no post-fix verification checkpoint — e.g., no instruction to re-run scripts/audit.ts to confirm DEAD_LINK/INDEX issues are resolved. Per the rubric, batch operations without validation cap workflow clarity at 3; the cap applies here. Not 2 — the sequence itself is thorough with explicit checklists.

Progressive Disclosure

A single well-sectioned SKILL.md with one bundled script (scripts/audit.ts) referenced by exact path and verified to exist and match its documented behavior. Structure is good and navigation is easy. Not anchor 5 — the issue-kind documentation and full report template sit inline, making the body denser than a pure overview with split-out detail; not anchor 3 — references that exist are clearly signaled and content is appropriately placed for a ~110-line skill.

Specificity

Names the domain (memex audit) and several concrete actions ('Runs a structural check, layers LLM-judgement findings, auto-fixes trivia with `--fix`') plus enumerated issue types ('orphans, dead SALP REFs, missing sections, contradictions or index drift'). It falls short of anchor 5 because 'structural check' and 'LLM-judgement findings' are umbrella labels rather than fully concrete actions, but it clearly exceeds anchor 3's 1-2 actions.

Completeness

Explicitly answers both: 'what' via 'Runs a structural check, layers LLM-judgement findings, auto-fixes trivia with `--fix`' and 'when' via the concrete trigger phrase 'Use when the user asks to audit a memex... for orphans, dead SALP REFs, missing sections, contradictions or index drift'. This matches the anchor-5 example pattern exactly.

Trigger Term Quality

'Use when the user asks to audit a memex' plus 'orphans', 'contradictions', 'dead SALP REFs' give good natural keyword coverage — a user would plausibly say 'audit my memex'. Not anchor 5 because common synonyms like 'health check' or 'dead links' are absent; not anchor 3 because the core natural terms are present and specific.

Distinctiveness Conflict Risk

'memex (long-term knowledge bank for AI agents)' and 'SALP REFs' define a clear niche with distinct trigger terms ('audit a memex', 'index drift') that no generic skill would claim. Minimal conflict risk, matching anchor 5.