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 exceptionally operational skill body: concrete read-only commands with exact pass/fail criteria for every category, a verification workflow with explicit refutation, severity calibration, and a validate-until-passing feedback loop, and a clean one-level-deep reference bundle that all resolves. The only real cost is efficiency — version-sensitive facts (October 2026 sweep dates, CVE-2026-66066, libvips and CLI versions) sit inline instead of a quarantined dated section, which will drift and adds tokens.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is command-dense and assumes Claude's competence — no paragraphs explaining what IAM, S3, or Vercel are, and nearly every clause carries operational content (exact commands, exact check criteria, verdict definitions). It falls short of anchor 5 mainly on the time-sensitivity guideline: specific dates and versions are inline rather than quarantined ("a real fleet sweep in October 2026" twice, "CVE-2026-66066 ... needs libvips ≥ 8.13; Heroku-24 ships 8.15.1", "Heroku CLI 8.x can't run pg:backups on Essential plans") plus a few scene-setting sentences ("Many small teams have no AWS account at all; production is a Vercel team, a few Heroku apps, and a Render service") that could be trimmed. These are minor, localized instances, so anchor 4 ('efficient; minor instances that could be trimmed') fits better than anchor 3's broader unnecessary-explanation pattern. | 4 / 5 |
Actionability | Fully executable, copy-paste-ready read-only commands throughout: `aws iam get-account-summary`, `gcloud projects get-iam-policy $PROJECT_ID`, `vercel api "/v6/deployments?..." | jq '...'`, `heroku pg:backups:schedules -a <app>`, `gh api repos/<o>/<r>/vulnerability-alerts`, `curl -s -o /dev/null -w '%{http_code}'`, and even a temp-file one-liner for comparing preview/production database hosts without exposing secrets. Every command block is followed by exact pass/fail criteria ("source is git (not `cli`), the deployed commit equals the production branch HEAD"), and validation is runnable via `node references/validate-findings.mjs findings.json`. Placeholders like `<team>` and `<app>` are standard parameterization covering the common cases, matching anchor 5. | 5 / 5 |
Workflow Clarity | A clear end-to-end sequence — Authorization → Scope the Audit → Audit Categories → "Verify, then record" (7 numbered steps: evidence first → refutation pass → verdict → severity → fingerprint → ext fields → write & validate) → Output Format → Boundaries — with explicit validation checkpoints and a feedback loop: "Run the refutation pass ... on every candidate before it is confirmed" and "write `findings.json`, and fix until it passes." Re-run recovery is specified ("A `rejected` record stays rejected unless the control it names changed"), and the verdict table and coverage table act as checklists. All commands are read-only with writes routed to the owner ("it's a write, so it prompts"), so the destructive-operation cap does not apply. This matches anchor 5's explicit-validation-with-feedback-loops pattern; anchor 4 would require a missing checkpoint, and none is missing. | 5 / 5 |
Progressive Disclosure | Scored against the actual bundle: all 7 files in `references/` (findings-format.md, findings.schema.json, paas.md, severity-calibration.md, to-sarif.mjs, validate-findings.mjs, verification.md) are referenced from the body with a stated purpose ("Netlify, Railway, and Fly equivalents are in [references/paas.md]"; "Calibrate severity with [references/severity-calibration.md]"), every link resolves, and references are one level deep — cross-links between reference files are lateral siblings, not nested depth. The split is deliberate and signaled: the common-case Vercel/Heroku checks stay inline because the finding-type definition and severity depend on them, while the other platforms and the shared findings workflow live in references. This matches anchor 5 (well-signaled one-level-deep references, appropriate split, easy navigation); anchor 4's 'minor organization gaps' would require an unsignaled or misplaced reference, and none exists. | 5 / 5 |
Total | 19 / 20 Passed |