The UI review skill with 587,920 installs (August 29, 2026, Skillselion catalog) does not contain a single accessibility rule. web-design-guidelines, Vercel's UI review skill, is 39 lines including frontmatter, and every rule it checks lives in a second repository it fetches at review time, from a URL that tracks main.
We run Skillselion, a directory of Claude Code, Codex and Cursor extensions ranked by real installs, and we read the skill file at commit b8caa26 of vercel-labs/agent-skills for our deep dive on the skill. Here is what the architecture actually implies for your reviews.
The file is an instruction to go get instructions
One sentence defines the whole design:
Use WebFetch to retrieve the latest rules. The fetched content contains all the rules and output format instructions.
Both halves of a review live somewhere else: what to check, and how to print what you find. The repo's README advertises "100+ rules covering accessibility, performance, and UX", and that number appears nowhere in the skill file. It arrives by download when the review runs.
That means the skill needs the network, and it asks for it by naming WebFetch, a Claude Code tool. In a sandboxed CI runner, or under an agent with no equivalent fetch tool, the skill still loads and still triggers, with nothing to check your code against.
The rules changed on August 18 and your install did not
The fetch target is raw.githubusercontent.com/vercel-labs/web-interface-guidelines/main/command.md. main is a branch. GitHub's history for that path shows five commits, three of them dated August 18, 2026, adding accessibility and media rules and revising the curly-quote guidance.
So a component that came back clean in July can come back with findings in August, and the diff that caused it sits in vercel-labs/web-interface-guidelines, a repository you do not depend on and were never asked to watch. Your rules improve when Vercel improves them, which is the defensible intent. The bill arrives as reproducibility: two reviews of unchanged code, weeks apart, can legitimately disagree.
Teams that need a review to hold still have the usual escape hatch: vendor the rules file into your own repo and point the agent at your copy.
The sheet it fetches is missing the rules with numbers in them
Vercel publishes three rule documents in the guidelines repo. Read on August 20, 2026: the README carries 129 rule bullets, AGENTS.md carries 107 written as MUST, SHOULD and NEVER, and command.md, the one this skill hard-codes, carries 103.
The coverage gap is not random. Under "Targets & Input", AGENTS.md holds two MUST rules the command sheet has no equivalent of: a hit target of at least 24px (44px on mobile), and a mobile input font size of at least 16px so iOS Safari stops zooming on focus. The command sheet also drops focus management, tooltip timing and most of the README's Design section.
Ask this skill "are my tap targets big enough?" and you get a guess. The fetched sheet carries no minimum hit-target rule, and nothing in the report tells you the question was out of scope.
What it does cover is worth having: ARIA labeling and semantic elements, focus-visible rings, form autocomplete, compositor-friendly animation, Intl formatting, hydration-safe inputs. The gap is a subset problem, not a quality problem.
It is a template, and the copies inherit Vercel's edits
The same repo ships writing-guidelines, the prose review skill: the same file with five lines changed (name, description, heading, summary, URL). Any raw markdown URL can become a review skill this way, and every skill built this way needs the network at review time and tracks a branch somebody else controls.
Our Agent Skill Clone Census, frozen at August 12, 2026, ranks web-design-guidelines the second most-copied original in the catalog: 36 identical or near-identical copies, behind Anthropic's frontend-design at 59. Each copy inherits the one URL, which points into a repository the copier does not control. The copies that keep that URL fetch Vercel's rules on every run and inherit Vercel's edits, which is an unusual property for a fork to have.
Contrast the sibling skill in the same repo: vercel-react-best-practices, the React performance skill, ships 70 rule files on disk after install. Same author, opposite architecture.
Should you install it?
Install it if you review UI in Claude Code and want the check driven by a published external standard instead of whatever the model feels strongly about that session. The rules on the other end are genuinely good, and the cost of keeping them current is one fetch per review.
Skip it if your agent runs without network access, or if your reviews need to be reproducible. Vendor the sheet instead.
If you want the generation-side counterpart, the guidelines repo pairs this skill with its AGENTS.md on purpose: the skill reviews what exists, AGENTS.md shapes what gets generated. Installing the skill does not put AGENTS.md in your project.
For the taste-layer alternatives developers actually cross-shop, see our hub of frontend design skills: Anthropic's frontend-design generates UI under a design brief, impeccable applies Paul Bakaus's design playbooks, ui-ux-pro-max looks up palettes and WCAG rows, and design-taste-frontend is the landing-page taste layer. web-design-guidelines is the only one of the five that generates nothing: it grades the files you already have.
Every quote above is verified against the repository at the commit named, section by section, in the full deep dive. The dive series index is at Deep Dives, and the full catalog at the skills directory.
Disclosure: we run Skillselion. Counts come from our live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs.
Top comments (0)