When users say "Review my interface" or "Check accessibility," the skill first reads the bundled references/command.md (a snapshot of Vercel's web interface guidelines), then reads the user-specified files or patterns, checks each rule one by one, and lists issues in a concise "file:line number" format; if no files are specified, it asks which ones to review. Version 1.1.0 of SKILL.md specifies the commit and SHA-256 of the snapshot source and requires that rules not be fetched from the network.
Why It Matters The rules are now bundled with the skill, pinned to a specific upstream commit, and include checksums, so reviews no longer require fetching an unlocked version of the rules from the internet each time.
skills.sh installs 890302 (+3400 / 24h)
Security Audit: Risk warnings present
- Initiates network requests or calls external APIs / remote MCPs
- Audit scope: GitHub vercel-labs/agent-skills/skills/web-design-guidelines/SKILL.md (scans only the SKILL.md text)
Rationale
- skills.sh installs 3400 in 24 hours, with a cumulative total of 890302 (indicating attention level); official Vercel publisher
- Pure documentation-type skill; the review process reads only local files and does not require accounts or keys
Usage
- Run npx skills add https://github.com/vercel-labs/agent-skills --skill web-design-guidelines, then tell the agent "Review the components under src/components according to the web interface guidelines"
Risks
- The audit warning about network requests is consistent with its purpose: it only appears in the maintenance step for "updating rule snapshots" (using gh and curl to fetch new versions); daily reviews do not connect to the internet
- The rules cover only items explicitly stated in the guidelines; review results still require human judgment for prioritization