Content
92%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
An expert-level audit skill body: actionable grep patterns, exact evidence commands, a verdict system with a refutation gate, an explicit validate-and-fix loop, and a report template rendered from a validated findings.json. The single weakness is token efficiency — regulatory reference material (the 18-identifier list, HITECH history, breach-timing table) is inlined where a tighter summary with a pointer would do.
Suggestions
Move the full 18-identifier enumeration and the breach-notification timing table into a reference file (e.g., references/phi-identifiers.md, references/breach-timing.md), keeping only the Safe-Harbor caveat and the 60-day headline inline with links.
Compress the HITECH section to its single engineering takeaway (BAs are directly liable for Security Rule violations) plus the penalty cap, dropping the historical framing.
Rewrite the vague audit-grep entries as literal patterns — 'search for session config without explicit timeout' should become concrete regexes for common session stores (e.g., express, django, spring) so every grep pattern is runnable as written.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with domain-specific material Claude cannot be assumed to know (section cites like §164.312(a)(2)(iii), addressable-vs-required semantics, ZIP-population thresholds for identifiers, vendor BAA stances), with no explanation of basic concepts — matching 'Efficient; minor instances of over-explanation that could be trimmed.' It is not 5 because the full 18-identifier enumeration, the HITECH background bullets, and the breach-timing table are regulatory lookup material that largely restates known facts or could be compressed into a reference file. It is not 3 because nothing is padded or pitched at a beginner. | 4 / 5 |
Actionability | Concrete, copy-paste-ready guidance throughout: literal grep patterns ('SELECT \\*.*patient'), an exact CLI command ('aws rds describe-db-instances --query … --output text'), example evidence records with code ('analytics.track("appointment_booked", { email, diagnosis })'), a fingerprint grammar with worked examples, an ext-field table, and runnable validation/export commands ('node references/validate-findings.mjs findings.json'). This matches the 'fully executable, copy-paste ready, covers the common cases' anchor; even the one vague pattern ('search for session config without explicit timeout') is offset by complete examples elsewhere. | 5 / 5 |
Workflow Clarity | The sequence is explicit and gated: scope/authorization → audit per safeguard category with starting hypotheses → close every candidate with one verdict 'and run the refutation pass … before anything is confirmed' → numbered steps (write findings.json sorted by fingerprint, 'Validate it and fix every error,' optional SARIF export) → render the report from the JSON. That is an explicit validation checkpoint with a fix-and-retry feedback loop, plus checklists (Scope, Breach response readiness) — matching the top anchor. Not 4, because there is no missing checkpoint: the validate-until-pass loop is stated, not implied. | 5 / 5 |
Progressive Disclosure | The SKILL.md body keeps the domain knowledge (rules, findings patterns, output template) inline and pushes the process machinery one level deep into real, clearly signaled files — [references/findings-format.md], [references/verification.md], [references/severity-calibration.md], [references/findings.schema.json], [references/validate-findings.mjs], [references/to-sarif.mjs] — all of which exist in the bundle and contain substantive content. No nested second-level references, no buried pointers; this matches the 'clear overview with well-signaled one-level-deep references' anchor. | 5 / 5 |
Total | 19 / 20 Passed |