Shaduf.
AI Web Design & Frontend Skills Catalog/Initial skill radar: create, review, then optimize

Initial skill radar: create, review, then optimize

Question: Which ready-made AI agent skills are most useful for web design and frontend development, and when should you use each one?

Answer: Start with a visual-direction skill when a new interface needs deliberate choices, add a UI-review skill when code already exists, and use a React/Next performance skill only when that stack is in scope. Those are complementary workflows, not a universal ranking.

What changed

We added three source-pinned listings: Anthropic Frontend Design at commit 41bbe19; Vercel Web Design Guidelines 1.0.0 at Agent Skills commit 063bee94; and Vercel React Best Practices 1.0.0 at the same release commit.

The practical split

  • Frontend Design: a strong stated fit for a landing page, dashboard, or substantial visual rework. It asks the agent to ground the design in the brief, make a plan, and critique generic defaults.
  • Web Design Guidelines: a conditional fit for a review pass. It fetches current Vercel rules, reads named files, and produces file-and-line findings. The fetched rules are mutable, so the package revision alone is not the whole dependency.
  • Vercel React Best Practices: a conditional fit for React/Next performance work. Its stated scope includes bundle size, waterfalls, rendering, and re-renders; it is not a general visual-design system.

What the sources do—and do not—prove

Source files prove what their authors publish: instructions, metadata, scoped tasks, and licences. They do not prove a particular agent will follow those instructions, that an interface will look good, or that a review catches every issue. We did not run a shared task or publish a performance comparison in this report.

Recommendation

Do not ask a skill to invent your product and call the result design. Give the agent your component inventory, key states, and a visual target; then use browser screenshots as acceptance evidence. Community discussion points toward that loop, but it remains a workflow hypothesis until we test it.

Sources

Search published pools, pages, reports, and evidence.