Conciseness

The body is dense and mostly earns its tokens — working quick-start code, a tight 5-step workflow, and high-value gotchas ("Never log the leaked value", "Checksum-gate card numbers", "Surrogates are not leaks"). A few rhetorical sentences could be trimmed without loss ("De-identification is verified, not assumed" framing, "It is the verification half of OpenMed's leakage-first ethos — gate on leakage, not on F1", which repeats the description), so it lands on anchor 4 (efficient, minor over-explanation) rather than anchor 5 (every token earns its place). It is clearly not anchor 3: nothing explains concepts Claude already knows.

Actionability

The quick start is fully executable, not pseudocode: a real Luhn implementation, a concrete DETECTORS regex dict, an actual API call (`openmed.extract_pii(deid_text)`) with entity handling, and a hard gate (`assert not leaked, "Release BLOCKED..."`). It is copy-paste ready and covers the common case end-to-end, matching anchor 5 rather than anchor 4 ('minor gaps').

Workflow Clarity

The 5-step workflow is clearly sequenced with an explicit validation checkpoint — step 5 "Block on any leak. The gate is binary for release: if findings is non-empty at high/critical, fail the export" — so the batch-operation cap of 3 does not apply. It falls just short of anchor 5 because the error-recovery loop is only implied ("so a reviewer can locate and re-redact") rather than an explicit re-redact-then-re-scan-until-clean step, and it is well above anchor 3, whose validation is missing or implicit.

Progressive Disclosure

No bundle files exist (no references/, scripts/, or assets/), and none are referenced from the body, so there are no broken or buried reference paths. The ~130-line body is well-sectioned (When to use, Quick start, Workflow, Hand-off, Edge cases, Standards) with everything reasonably kept inline for a single-purpose skill and external links grouped in a dedicated Standards section. This sits at anchor 4 (good structure, content appropriately placed) rather than anchor 5 because there are no well-signaled one-level-deep references at all, and at ~130 lines a detector-pattern table or standards list could plausibly live in a reference file.

Specificity

The description lists multiple concrete actions — "Adversarially scan already-de-identified clinical text for residual identifiers and emit a leakage report that blocks release on any hit" plus "format and checksum detectors (SSN, Luhn for card numbers, MRN/account patterns, emails, phones, dates), entropy heuristics... severity scoring, and a hard block-on-leak rule" — with comprehensive capability coverage. It exceeds anchor 4 ('minor gaps in coverage') because every core capability of the skill is named explicitly, so it is not the level below.

Completeness

Both questions are answered explicitly: what — "scan already-de-identified clinical text for residual identifiers and emit a leakage report that blocks release on any hit" with a Covers-list of detector families; when — "Use after OpenMed de-identification when the user asks to verify a redaction, prove no PHI/PII leaked, gate a dataset before sharing, or run a second-pass detector." This matches anchor 5 (explicit what AND when with concrete trigger phrases) and clearly exceeds anchor 4, whose 'when' is weaker ('Use when working with PDF files').

Trigger Term Quality

Natural user phrasings are comprehensively covered: "verify a redaction", "prove no PHI/PII leaked", "gate a dataset before sharing", "run a second-pass detector", with synonyms (PHI/PII, redaction, leak, gate, de-identification). A user needing this skill would naturally say several of these; anchor 4 ('a few natural terms missing') does not fit because both the synonyms and the colloquial trigger phrases are present.

Distinctiveness Conflict Risk

It carves out a clear niche — auditing the *output* of de-identification rather than de-identifying — and explicitly disambiguates from the sibling skill: "Pairs after deidentifying-clinical-text" and "This is the verification half of OpenMed's leakage-first ethos." The trigger conditions (verify/gate/second-pass on already-de-id text) would not plausibly fire the de-id skill itself, so conflict risk is minimal (anchor 5).