Content
85%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 actionable, well-sequenced audit protocol with excellent progressive disclosure — every phase has concrete commands, templates, thresholds, and validation gates, and all 22 reference files are real and one level deep. The main weakness is token efficiency: gate and Incomplete criteria are repeated across up to four sections, and the closing reference table duplicates the inline links, so the ~470-line body could be tightened by roughly a quarter without losing guidance.
Suggestions
State the Incomplete-criteria thresholds (0.5s median manifest gap, Phase-3 minimum duration, screenshots/console-reads per route) once in the Hard gates or audit-the-audit table and cross-reference them from the Verdict states section and verdict block template, instead of repeating them in four places.
Move rationale/editorial asides (e.g., "CWV-strict is for landing pages and marketing teams...", "A fresh sub-agent works because the original drafter is invested in its own output") into the reference files that already carry the full protocols, keeping the body to the rule itself.
Drop either the bottom Reference files table or the inline reference links — keeping one navigation mechanism instead of two would cut ~25 lines with no loss of discoverability.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is operational rather than educational (no basic-concept teaching), but key thresholds are structurally redundant: the < 0.5s manifest-gap and Phase-3 timing rules each appear in four places ("manifest timestamps clustered < 0.5s apart", "Manifest median gap between entries < 0.5s", "Median gap between manifest entries < 0.5s", "median gap [N]s (< 0.5s = Incomplete)"), and the bottom Reference files table restates nearly every inline link. This fits anchor 3 (mostly efficient but could be tightened) rather than 4, where redundancy would be only minor; it is above anchor 2 because no section is padded with explanations Claude doesn't need. | 3 / 5 |
Actionability | Guidance is copy-paste ready throughout: exact fallback chains ("`.jez/audit-config.yml` → `audit-config.yml` → `.audit/config.yml`"), a tool-selection table with stop conditions, real selectors in the manifest template ("textarea[placeholder*=\"message\"]", "li[data-mention-id=\"assistant\"]"), port mappings ("5173 Vite, 3000 Next/Rails, 8000 Django/Laravel"), the literal sub-agent self-critique prompt, and a fill-in verdict block template. Fully executable with specific examples covering the common cases. | 5 / 5 |
Workflow Clarity | Seven phases are explicitly ordered with validation checkpoints everywhere: pre-flight "Stop if any fail" gates and capability tests before starting, per-action verification in the manifest ("Verified input cleared within 1000ms"), the audit-the-audit meta-check, and a fix-and-verify phase that re-walks the affected slice. Clear sequence, explicit validation steps, feedback loops, and checklists — the anchor-5 pattern; it does not miss the anchor-4 weaknesses of implicit checkpoints. | 5 / 5 |
Progressive Disclosure | All 22 referenced files exist, are one level deep, and are signposted inline at the point of need plus a When→Read navigation table at the bottom; the SKILL.md itself stays an overview while full protocols live in references. Cross-links between reference files (e.g., scenario-tests.md → round-trip-workflows.md) are lateral see-also pointers, not chained deferral, so navigation remains easy — anchor 5. | 5 / 5 |
Total | 18 / 20 Passed |