Content
57%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.
The body is a dense, prescriptive review protocol with a concrete severity taxonomy, required finding fields, and deterministic decision rules, and it wisely defers full protocols to one-level-deep references. However, the multi-phase workflow order must be inferred from scattered bullets, three sections repeat the same anti-style guidance, and half the reference bundle (lenses.md, report.md, request-template.md) is unreachable from SKILL.md.
Suggestions
Collapse 'Red Flags', 'Rationalization Prevention', and 'Anti-Patterns' into a single section — the 'ignore style, focus on behavior' instruction currently appears four times across Principles, Red Flags, Rationalization Prevention, and Anti-Patterns.
Replace the scattered phase rules in 'Review Principles' with a numbered phase sequence (deterministic gate → blind review → freeze findings/SHA-256 → reconcile/resume) so the workflow order is explicit rather than inferred.
Add links to references/lenses.md, references/report.md, and references/request-template.md from SKILL.md (or from orchestration.md), and fix lenses.md's dangling reference to '../../common-security-audit/references/signals.md'.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and directive with no padding explaining known concepts, but the 'ignore style / focus on behavior' instruction is repeated four times ('do not nitpick' in Principles, 'Stop if you are reviewing style only' in Red Flags, "'Only style matters here'" in Rationalization Prevention, and 'No Nitpicking' in Anti-Patterns), and the inline 'Review Checklist (Mandatory)' duplicates references/checklist.md. This fits anchor 3 ('could be tightened'); not 2 because there is no unnecessary conceptual explanation, not 4 because the redundancy goes beyond minor trimming. | 3 / 5 |
Actionability | Concrete, executable guidance throughout: severity taxonomy 'Use only `[Blocker]`, `[Major]`, or `[Nit]`', required finding fields with an enum ('Prevent-by: lint | type | test | eval | runtime | guidance | none'), and deterministic decision rules ('a current orchestrator-observed required-check `FAIL` returns `CHANGES REQUESTED`', 'If resume is unavailable, BLOCKED'). Not 5: no worked example of an actual finding appears in the body (deferred to output-format.md), and governance terms like 'orchestrator-observed check receipt' are underspecified in-body. Not 3: the guidance is executable instruction, not pseudocode. | 4 / 5 |
Workflow Clarity | A multi-phase sequence exists — deterministic gate, 'phase one receives only the brief, cumulative diff', 'Freeze findings and SHA-256 before author claims', 'Reconcile: resume the same reviewer with frozen findings' — but it is scattered across prose bullets with no ordered steps; the reader must infer the phase order. Checkpoints are named but implicit in sequence, matching anchor 3; not 2 because sequence and validation are present, not 4 because the order is not clearly laid out. | 3 / 5 |
Progressive Disclosure | The four in-body references are clearly signaled with purpose and all exist ('For complete phases and formats, load [Adversarial Review Protocol]', 'Load [orchestration] when ticket/AC, PR metadata, or specialist lenses are needed'). However, scoring against the actual bundle: three of seven reference files (lenses.md, report.md, request-template.md) are not linked from SKILL.md or from any other reference file, and lenses.md itself contains a dangling link to '../../common-security-audit/references/signals.md'. Half the bundle being undiscoverable exceeds the 'minor organization gaps' of anchor 4, fitting anchor 3. | 3 / 5 |
Total | 13 / 20 Passed |