Content
77%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 unusually actionable, well-sequenced design skill with real install commands, executable code skeletons, and a mechanical pre-flight checklist — workflow clarity and actionability are top-tier. Its weaknesses are structural: a monolithic ~1100-line body with material that should be split into reference files, and meaningful duplication between Sections 4, 9, and 14 that inflates token cost without adding new guidance.
Suggestions
Move the Section 10 pattern vocabulary, Section 9.F tell catalog, and Appendices B/C into reference files (e.g. references/patterns.md, references/ai-tells.md, references/design-systems.md) linked one level deep, leaving SKILL.md as the overview plus core dials and workflow.
Cut the duplication between Section 4.7, Section 9.F, and the Section 14 checklist — the pre-flight checklist can reference section numbers ("Eyebrow count check per 4.7") instead of restating each rule with examples.
Trim passages Claude already knows (breakpoint tables, Lighthouse/CWV targets, "never transition: all") to one-line reminders, keeping the token budget for the non-obvious taste rules that are this skill's actual value.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most of the ~1100-line body is dense, original, curated rules Claude does not already know (dial tables, AI-tell catalog, layout caps), so it is not the anchor-2 pattern of explaining basics. However, it could be tightened: the Pre-Flight Check (Section 14) restates nearly every Section 4/9 rule a second or third time, Section 9.F overlaps Section 4.7 heavily, and a few passages cover known ground (standard Tailwind breakpoints, "LCP < 2.5s ... run Lighthouse", "never transition: all"). | 3 / 5 |
Actionability | Fully executable throughout: exact Tailwind class recipes (`min-h-[100dvh]`, `text-4xl md:text-6xl tracking-tighter leading-none`), real install commands (Appendix A: `npm install @carbon/react @carbon/styles`, `npx shadcn@latest init`), three complete copy-paste GSAP/Motion component skeletons (5.A/5.B/5.C), concrete placeholder URL patterns (`https://picsum.photos/seed/{descriptive-seed}/{w}/{h}`), and canonical doc links. Specific examples cover the common cases with explicit failure modes and fixes. | 5 / 5 |
Workflow Clarity | The workflow is clearly sequenced — brief inference (0) → dials (1) → design-system selection (2) → architecture (3) → directives (4-5) → guardrails (6-8) → redesign protocol (11) → final Pre-Flight Check (14) — with explicit validation checkpoints: mandatory dependency verification before any import, redesign audit before changes, copy self-audit before ship, and a mechanical final checklist with countable pass/fail criteria (e.g. eyebrow count ≤ ceil(sectionCount/3)). Error-recovery feedback loops are present ("If validation fails: review, fix, run again" style guidance in the dial-override and audit sections). | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are all absent), so everything lives inline in a single ~1100-line file. Section headers and numbered cross-references give it real structure, so it is above anchor 2, but large bodies that clearly belong in separate reference files are inlined: the Section 10 pattern vocabulary, the Section 9.F tell catalog, Appendix B's URL lists, and Appendix C's full Liquid Glass CSS block. Per the rubric's own definition ("vendored reference content"), these are prime candidates for one-level-deep reference files. | 3 / 5 |
Total | 16 / 20 Passed |