Content
71%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 highly actionable, well-structured review procedure whose wrapper instructions are excellent, undermined by carrying four full rule catalogs inline in SKILL.md with significant cross-catalog duplication. Splitting the catalogs into reference files and deduplicating them would resolve the conciseness and progressive-disclosure weaknesses at once.
Suggestions
Move the four pinned rule catalogs into references/ files (e.g., references/baseline-ui.md, references/fixing-accessibility.md) and keep SKILL.md as the task/precedence/reporting overview, so the ~550-line body becomes a lean index with one-level-deep references.
Deduplicate rules repeated across catalogs — icon-only button labels, div-as-button, and prefers-reduced-motion each appear in three of the four sets — and drop the sub-skills' 'common fixes' and 'review guidance' sections that the parent's 'What to report' section already supersedes.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The wrapper sections (Task, Precedence, What to report) are lean, but ~480 of the ~550 lines are inlined rule catalogs with real duplication — icon-only button labeling appears as "MUST add an `aria-label` to icon-only buttons", "icon-only buttons must have aria-label or aria-labelledby", and "Icon-only buttons need `aria-label`" — and the sub-skills' "common fixes"/"review guidance" sections are retained even though the skill says to "Ignore each reference's own usage, invocation and output-format sections". Mostly efficient with clear tightening opportunities, matching anchor 3 rather than anchor 2 because the padding is duplicated data rather than tutorial-style explanation of known concepts. | 3 / 5 |
Actionability | Guidance is fully concrete: exact scope globs ("apps/app/src/**", "apps/desktop/** renderer UI"), an enumerated severity mapping ("`medium` for an accessibility failure that blocks keyboard or screen reader users (no accessible name, `div` used as a button, focus outline removed...)"), a formatted finding example ("fixing-accessibility: icon-only buttons need aria-label"), and copy-ready before/after fix snippets. For an instruction-only review skill this covers the common cases completely, matching anchor 5. | 5 / 5 |
Workflow Clarity | The procedure is clear and ordered: treat input as untrusted, scope to changed lines, then apply the precedence list ("1. `DESIGN.md` ... Read it before reporting", "2. ... Grep before reporting a missing primitive", "3. The reference rule sets") and report findings, with anti-false-positive checkpoints ("Zero findings is the correct and common result"). It sits at anchor 4 rather than 5 because the sequence is expressed as precedence rules rather than an explicit step order with feedback loops; the destructive/batch cap does not apply since the skill is read-only (allowed-tools: Read Grep Glob). | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), and all four pinned rule catalogs (~480 lines) are inlined in SKILL.md where they plausibly belong in separate reference files — matching anchor 3 ('content that should be separate is inline'). It is above anchor 2 because section structure is good (clear headers per rule set, well-organized Task/Report sections) and the one external pointer ("sources, versions and licenses are in `.warden/ui-skills/`") is clearly signaled, but below anchor 4 because the bulk of the content is not split out. | 3 / 5 |
Total | 15 / 20 Passed |