Lighthouse Optimizer

Lighthouse reports symptoms, often many per cause. Treating the failing list as N independent fixes is the slow, wrong path — it churns on duplicates and sometimes regresses the score. One oversized hero image fails LCP, properly-size, next-gen formats, defer-offscreen, and contributes to CLS. That's one fix, not five. A render-blocking head fails three audits and pushes LCP. CLS almost always has two causes at once — image dimensions and font swap reflow. The skill clusters correlated audits to root causes first, names which audits each cause explains, and orders the work by how few changes clear the most points.

Then the actual fixes — concrete, not categorical. For accessibility, you get the hex code and the contrast ratio it achieves, not "darken the text". For LCP, the skill checks whether the element is an image (preload, fetchpriority, drop lazy) or post-hydration text (no image trick helps — the structural fix is SSR/SSG or prerender). For CLS, both common causes get addressed because fixing one rarely clears it. Third-party scripts get the facade pattern, defer, consent-gating, or Partytown — and the skill is explicit that vendor payload weight is not under your control.

The honesty layer is the part most score-optimisers skip. Lighthouse is a single throttled lab run; what users feel is field data (CrUX, real-user CWV). A 100 in lab can coexist with a site that fails real users — Lighthouse only catches about a third of WCAG criteria, so a green accessibility score is not "accessible". When significant weight sits in vendor scripts, the achievable lab score is capped and not stably 100; the skill says so plainly instead of promising a number it can't deliver.

It also refuses harmful fixes outright. Hiding a consent banner, swapping the LCP element for a placeholder, gutting content for a metric — these trade UX, legal compliance, or content for a lab integer and don't count as fixes even when the user says "do whatever it takes".