cg-web-lighthouse-optimizer

Hiding content from the audit, lazy-loading the LCP image, preloading everything, removing interactive elements to pass accessibility checks, keyword stuffing for SEO — the skill lists these tricks and forbids them. Mobile and desktop are analyzed as separate test conditions, and a lab score is never mistaken for what real users experience.

A single report, a mobile/desktop pair, an array, a PageSpeed Insights wrapper or JSON pasted as text: Claude normalizes it, checks for runtime errors, warnings, mismatched versions or URLs, and only then decides what to optimize.

The audit title is only a pointer. Claude reads the details — savings, affected URLs, nodes, timings — and follows them into the codebase: the actual LCP element, the dependency behind the long tasks, the unsized image behind the layout shift, the font, the third-party script, the cache header.

Audit IDs change between major releases, and Lighthouse 13 folded many older performance audits into shared insights. Claude reads the report's own version and weights, and maps old and new audit IDs to the same problem domain instead of fixing audits that don't exist in the report.

A change that probably helps is not a measured improvement. When Lighthouse cannot be run again, the skill reports what it changed and why — and nothing more:

Not score → hack → score. One root-cause group at a time, validated with lint, build and tests, then measured again on mobile and desktop before the next change.