Content
61%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.
A dense, disciplined instruction skill with a concrete output contract and strong progressive disclosure from the entry file. Its main weaknesses are triple-duplicated rules across the body and reference files, an implicit phase sequence that requires assembly from principle bullets, and orphaned/broken-link bundle files.
Suggestions
Merge the overlapping 'Red Flags', 'Rationalization Prevention', and 'Anti-Patterns' sections into a single anti-patterns section — all three repeat the same no-praise / evidence-required / behavior-over-style rules.
Replace the body's duplicated deterministic-gate and finding-template text with a single pointer to references/output-format.md (the same contract currently appears verbatim in three places).
Add an explicit numbered phase sequence in the body (blind review → deterministic gate → freeze findings + SHA-256 → author claims → reconcile) instead of encoding the order inside prose principle bullets.
Link lenses.md, report.md, and request-template.md from SKILL.md or orchestration.md, and fix or remove the broken cross-skill links in lenses.md and report.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is terse but redundant: "Red Flags", "Rationalization Prevention", and "Anti-Patterns" each repeat the same three ideas (no praise, claims need evidence, behavior over style) — e.g., praise is banned in "Findings first; no strengths or praise section", "Stop if you are praising before reviewing", and the output-format rule. The deterministic-gate and finding-template text is also duplicated almost verbatim across the body, adversarial-review.md, and output-format.md, which could be consolidated to pointers. | 3 / 5 |
Actionability | The guidance is concrete and executable for an instruction-only skill: exact severity tags, required finding fields ("Evidence:", "Why:", "Fix:", "Prevent-by: lint | type | test | eval | runtime | guidance | none"), named defect classes ("No N+1 queries, memory leaks", "injection, secrets, auth leaks"), and a specific three-part test-evidence gate. It falls short of 5 only because no worked finding example appears inline (it is deferred to output-format.md). | 4 / 5 |
Workflow Clarity | The multi-phase process (blind first, deterministic gate, freeze findings and SHA-256, reconcile after author claims) is present and its checkpoints are explicit, but the sequence is embedded in dense principle bullets rather than an ordered workflow — the reader must assemble the phase order from prose, and error-recovery steps live only in the referenced adversarial-review.md. This matches the 'steps present but sequence/checkpoints implicit' anchor more than the clear-sequence 4 anchor. | 3 / 5 |
Progressive Disclosure | The body is a well-organized overview with clearly signaled, one-level-deep links to real files (adversarial-review.md, orchestration.md, checklist.md, output-format.md) plus purpose guidance for when to load each. The gap keeping it below 5: lenses.md, report.md, and request-template.md are orphaned (never linked from SKILL.md or any reference), and lenses.md and report.md contain links to non-existent cross-skill paths (../../common-security-audit/references/signals.md, ../../common-skill-creator/references/rubric.md). | 4 / 5 |
Total | 14 / 20 Passed |