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.
The body is an unusually well-crafted orchestration document: lean, mode-aware, with an explicit sequence and a genuine verification checkpoint before reporting. Its fatal flaw is the bundle itself — seven of the eight referenced rule and template files do not exist, so the progressive-disclosure structure it is built around cannot function in practice.
Suggestions
Add the missing bundle files (rules/user-path.md, rules/accessibility.md, rules/ux-writing.md, rules/dark-patterns.md, rules/visual-design.md, rules/verify-before-reporting.md, and templates/review-report.md) — every Pass 1/Pass 2 procedure, the copy and dark-pattern rules, and the report template currently resolve to dead links.
Resolve the `<base>` placeholder in the scoping command (e.g., "merge-base with the default branch") so `git diff --name-only <base>...HEAD` is executable without guesswork.
Provide a minimal inline fallback (severity labels plus required finding fields) for the report format when templates/review-report.md is absent, mirroring the existing graceful-degradation pattern for missing companion skills.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and dense — a mode table, terse numbered steps, and hard rules with no explanation of concepts Claude already knows (it never defines WCAG or React Native; it even delegates rationale to references/path-first.md, which notes "The base model already knows WCAG 2.2 success criteria unaided"). Not 4: no over-explanation was found that could be trimmed. | 5 / 5 |
Actionability | Concrete, executable guidance throughout: an exact scoping command (`git diff --name-only <base>...HEAD`), a severity-criteria table, per-mode work order, and specific directives like "Contrast computed from the actual tokens and resolved colors, in every theme" and "Report at most ten findings". Not 5: `<base>` is left unresolved, and the deepest procedures are delegated to rule files rather than given concrete inline steps. | 4 / 5 |
Workflow Clarity | The sequence is explicit per mode ("Pass 1 → Pass 2 → Verify → Report"), Pass 1 is a numbered checklist, and the Verify section is an explicit validation checkpoint with a feedback loop: "Every claimed behaviour cites the `file:line` that proves it" and unprovable claims are labeled unverified with "what would confirm it". Not 4: checkpoints are present and explicit, with error-recovery behavior even for missing companion skills. | 5 / 5 |
Progressive Disclosure | Scored against the actual bundle: the body references eight files, but only references/path-first.md exists — rules/user-path.md, rules/accessibility.md, rules/ux-writing.md, rules/dark-patterns.md, rules/visual-design.md, rules/verify-before-reporting.md, and templates/review-report.md are all missing, so the entire detail layer is unreachable. The signaling and one-level-deep structure are well designed, which keeps this above 2, but 7 of 8 broken references is far beyond the "minor organization gaps" of 4. | 3 / 5 |
Total | 17 / 20 Passed |