// the find
addyosmani/web-quality-skills
Agent Skills for optimizing web quality based on Lighthouse and Core Web Vitals.
A collection of markdown-based Agent Skills that tell coding agents (Claude Code, Codex, Gemini CLI, Cursor) how to audit and fix web performance, accessibility, and SEO issues using Lighthouse and Core Web Vitals. It's for developers who already have an AI coding agent and want it to run structured, evidence-based audits instead of guessing at fixes.
The measurement model is the best part: it explicitly separates CrUX field data, first-party RUM, DevTools lab traces, and static source inspection, and calls out that a single PerformanceObserver run in a browser is lab data, not real-user data. That's a distinction most 'AI audit my site' tooling gets wrong. Each skill keeps its SKILL.md under 500 lines and pushes detail into references/ (WCAG.md, LCP.md, STRUCTURED-DATA.md, etc.), which is the right call for keeping an agent's context usable. It also ships concrete numeric thresholds (LCP/INP/CLS cutoffs, JS/CSS/image budgets) instead of vague 'optimize your images' advice, and has fallback behavior when Chrome DevTools MCP isn't available rather than just failing.
This is markdown instructions, not code — despite the 'Shell' language tag, there's essentially one script (analyze.sh) and no automated way to verify the skills actually make an agent produce correct audits; quality is entirely dependent on how well the target LLM follows the prompt. There's no CI or test suite checking that the guidance stays accurate as Lighthouse/CWV scoring changes. Installation is fragmented across four different plugin ecosystems (npx skills, Claude Code plugin marketplace, Codex plugin marketplace, Gemini extensions) that are all young and could break independently. And because the value is just well-organized prompting, there's nothing stopping you from getting most of the benefit by reading the reference docs yourself instead of installing anything.