The Best Claude Code Design Skills, Actually Tested (2026)
Search “claude code design skills” and you get link dumps — twenty repos, a one-line description each, no indication of which ones actually do anything. So we scored them.
How we scored: 10 dimensions, deliberately not averaged
39 candidates found, 19 deep-reviewed against a 10-dimension rubric, 12 shortlisted. Every review was done by reading the SKILL.md, README and source — nothing was installed or executed, and where a candidate would need execution to judge fully, we say so rather than guessing.
One rule shaped everything below: stars and install counts are not a rubric dimension. They appear as unverified context or not at all. A skill with 4k stars and a single generic prompt scores below a 40-star repo with a working rule engine, and several did.
There is no single aggregate number, and that is the point. A skill can be excellent on domain knowledge and useless on validation — averaging those into one score hides exactly the thing you need to know before installing it.
- Design usefulness — does it improve output *quality*, or just speed?
- Implementation depth — a real pipeline, or prompt text with a skill wrapper?
- Design-domain knowledge — named, specific expertise with reasoning, or "make it look modern"?
- Reusability — retargetable to a new project, or hardcoded to one demo?
- Documentation quality — can a competent engineer predict its behaviour without running it?
- Example / demo quality — credible evidence of real output, or marketing renders?
- Validation quality — does it inspect and fix its own output, or generate and stop?
- Curify relevance — does the underlying pattern map onto a real production workflow?
- Social-media value — safe and useful to recommend to a general audience?
- Security / operational risk — what does it need access to?
The scale is anchored, not vibes. A 1 on design-domain knowledge is generic adjectives with no concrete rule. A 5 is typography-weight rules, platform compliance rules, cinematography vocabulary or WCAG criteria — traceable to a real discipline.
The shortlist: 12 that survived
Five S-tier, seven A-tier. Each one gets its caveat stated in the same breath as its strength, because the caveat is usually the thing that decides whether you can actually use it.
| Skill | Strongest capability | The caveat that matters |
|---|---|---|
| impeccable | 59 deterministic rules that run with no LLM call, plus 23 design commands across 14 agent tools | Needs an npm CLI installed; anti-pattern judgments go stale as trends shift |
| styleseed | Dual-gate validation — a 0–100 rule score and rendered-pixel verification before shipping | Tightly coupled to React + Tailwind + Radix; its +5.3-point benchmark is self-reported |
Anthropic frontend-design | The most complete single document on colour, type, motion, hierarchy and copy in the whole set | Self-critique is manual — a human or agent still has to look at the screenshot |
| product-shots | Platform compliance baked into the prompt, plus a 14-point identity lock across 9 angles | Requires a paid external image-generation API key to run at all |
| guizang-social-card-skill | A post-render validator that catches 9 classes of layout fault before the image ships | AGPL-3.0 — any derivative or SaaS use must open-source modifications |
| designer-skills | A /design-flow orchestrator chaining 8 skills with skip/confirm gates and persisted state | An unrelated repo shares the exact same name — always link the julianoczkowski URL |
| excalidraw-diagram-skill | The cleanest render → screenshot → inspect → fix loop in the entire review set | No license stated in the reviewed content |
| accesslint / skills | A 5-stage WCAG 2.2 lifecycle with honest severity *and confidence* grading | Full functionality needs an MCP server and Chrome CDP automation |
| web-design-guidelines | Grounded in Vercel's maintained guideline set, zero dependencies, agent-agnostic | Audits only — it flags, it does not fix or visually verify |
| Codex-Skills | ~70 skills, genuinely Codex-first by architecture rather than a Claude port | Benchmarks are self-reported on 2 runs, not independently reproduced |
| visual-verdict | A JSON verdict contract gated at 90+, built to drive an iterate-until-pass loop | No license stated; lives inside a larger personal collection |
| brand-skills | 15 composable skills persisted as a git-tracked brand.yaml, not chat output | No visual validation loop; its "70+ agents" figure is an unverified vendor number |
If you install exactly one and want the lowest possible setup cost, take web-design-guidelines. If you want the one most likely to change what your agent produces, take impeccable.
The property that separates a real skill from a prompt in a folder
After nineteen reviews, one thing predicted quality better than any other: does the workflow check its own output?
Generation and verification are different competencies. Most "design skills" conflate them — they try to generate well the first time and then stop. The strong ones separate the two steps and put a gate between them:
- impeccable encodes 59 known failure modes as deterministic rules. The detection step runs with no LLM call at all — same input, same flagged issues, every time.
- excalidraw-diagram-skill renders its own diagram with Playwright, screenshots it, detects overlap and misalignment, and fixes before delivery.
- guizang ships
validate-social-deck.mjs, which catches 9 classes of layout fault *after* render.
- styleseed gates on a rule-compliance score and rendered pixels — both, not either.
- visual-verdict is the cleanest form: a machine-checkable contract of
score,verdict,differences[]andsuggestions[], thresholded at 90+ specifically so a loop can iterate until it passes.
The insight generalises past coding agents. Known, recurring failure modes should be caught by cheap deterministic checks; model judgment should be reserved for things that genuinely require taste — not for catching "used the default font again."
If you are evaluating a design skill yourself, that is the question to ask first. Not how many rules it claims. Whether anything looks at the output.
What got cut, and why
Being reviewed and not shortlisted is a finding, not an omission. Four deep-reviewed candidates never made any cut:
- theme-factory — classified
thin_prompt_wrapper. Ten preset themes, no validation step, no colour-theory reasoning exposed to the user.
- hyperframes-creative — the SKILL.md 404'd. We fell back to the repo root and third-party skill indexes, flagged it as partial verification, and scored it lower for exactly that reason.
- platform-design-skills — the repo's own header and body disagree on its rule count (300+ versus 450+). Minor on its own, but we could not open the rule files to check either number.
- photo-abstract-editorial — no stated author and no license. Lowest trust of anything reviewed.
Three genuinely A-tier resources were also cut, for fit rather than quality, and are worth knowing about:
- webapp-testing — excellent, but it is QA infrastructure; it scores lowest of any shortlisted item on design-domain depth.
- figma-implement-design — a genuinely useful validation-checklist pattern, but its host repo is marked deprecated.
- video-shotcraft — 152 shot-recipe cards, the largest domain-knowledge library we found anywhere, and the highest social-media-value score of any candidate. Cut only for having no automated validation loop.
Two more results worth recording: one seed name (design-process-pack) could not be resolved to any real resource, and openai/plugins — the official repo — was found archived as of 2026-08-16.
What this review does not tell you
Every claim here comes from reading source, not running it. No third-party skill was installed, no script executed, no MCP server added, no credentials used.
That means behavioural claims — does the validation loop actually catch what it says it catches? — rest on a static read. Where a candidate would need execution to judge properly (running product-shots' image pipeline, or accesslint's Chrome-CDP automation), it is recorded as untested with the reason given, rather than scored on a guess.
We also deliberately did not deep-review 8 candidates whose sources we confirmed real but did not take through the full process. Their notes are verified-but-shallow, and should not be read as equivalent evidence.
Where this lands for us
We do the same work these skills do, on the generation side rather than the code side — which is why the validation-loop finding is the one we care about most.
- Brand Direction Explorer — one brief into multiple distinct creative directions, the same "explore before committing" step
designer-skillsenforces with its/design-flowgates.
- Character Sticker Sheet — a 9-pose expression sheet from one character, which is the identity-lock problem
product-shotssolves with 14 anchors across 9 angles.
- Die-Cut Sticker File — where a design stops being a picture and becomes a production file with a cut line, bleed and CMYK.
Browse more at design templates and branding.
The short version
Twelve are worth your time. Read the caveat before you install — guizang is AGPL-3.0, product-shots needs a paid API key, excalidraw-diagram-skill and visual-verdict state no license at all, and two strong-looking candidates could not be fully verified.
And when you evaluate the next one yourself, skip the star count and ask the single question that actually separated this field: does anything check the output before it ships?
Related Articles
design-branding
AI Packaging Design: From Brief to Mockup Set in 30 Minutes

Part 2: Agents vs Workflows – From Control to Intelligence

