Core Web Vitals Reports Tell You What. Not How.

Every site I have ever audited already has a Core Web Vitals report. Not one of them was short on dashboards. What they were short on was the next step.

The Core Web Vitals workflow most teams run is: open Search Console, see which URL groups fail, feel bad, close Search Console. The report tells you a page group is failing LCP. It does not tell you why, and the three metrics fail for reasons that live in three different parts of your stack. That translation gap — from a red number to a specific change in a specific file — is where the work actually is, and it is the part that agent tooling is unusually good at.

Three metrics, three different failure surfaces

This is the core of it, and it's why "improve your Core Web Vitals" as advice is worthless:

• LCP fails for server and delivery reasons: slow Time to First Byte, render-blocking resources in the head, an unoptimized hero image without priority hints, missing cache headers. The fixes live in infrastructure and templates.
• CLS fails for layout reasons: embeds and ads that inject late, images and iframes without reserved dimensions, web fonts that swap metrics on load. The fixes live in markup discipline.
• INP fails for compute reasons: long tasks on the main thread, hydration payloads, third-party scripts. The fixes live in code splitting and dependency audits.

Three metrics, three owners — an infra change, a markup change, a code change. A single failing page group can require a backend engineer, a template fix, and a bundle diet, and the report groups them under one red badge as if they were one problem.

What the skills contribute

In our directory, two skills cover the two halves of this translation, and they happen to be the two highest-scored entries in the category.

performance-monitor by aaron-he-zhu scores 8.9 and is the interpretation layer: it reads the vitals, classifies which metric is failing against which threshold on which page group, and — critically — distinguishes a real user (field) failure from a lab artifact. That last distinction matters more than people credit. A lab tool showing 2.8s LCP on a throttled connection and CrUX showing passing LCP at the 75th percentile is telling you the problem exists only for a user segment you may or may not care about. Interpreting which number governs is a real skill, and it is the first half of the translation.

technical-seo-checker, also by aaron-he-zhu, scores 9.3 with structure at 10/10 — the highest score in our SEO category, and the fastest-downloading SEO skill we list (⬇14). It is the structural layer: crawl paths, redirect chains, canonical consistency. Redirect chains are the crossover point — each hop in a chain is added latency on the critical path to the LCP resource, and the same structural sloppiness that wastes crawl budget also costs you time-to-first-byte. A chain of three 301s is simultaneously an SEO finding and an LCP finding, and this is where the two skills compose.

Run together, the agent workflow is: performance-monitor names the failing metric and page group → technical-seo-checker maps the structural causes on those URLs → the agent walks to the probable fix — the un-preloaded hero image, the chain, the unreserved iframe — and proposes the patch. Number, cause, file.

Why this is agent-shaped work

Look at what the human expert actually does with a failing INP: open DevTools, find the long task, trace it to a script, identify the script as a third-party widget or a hydration block, decide whether it can be deferred. Most of that is searching and tracing, not judgment. It is exactly the kind of work agents do well — bounded, evidence-driven, with a verifiable answer at the end.

The judgment calls remain human, and should: whether an embed is worth its CLS cost, whether a third-party script's business value justifies its main-thread tax. Those are tradeoffs, not lookups. The skill's job is to make sure the human is deciding the tradeoff instead of hunting for the cause.

The industry around Core Web Vitals sells dashboards, and has for years, which is why every team has five reports and no fixes. The scarce commodity was never measurement — the metrics are free. The scarce commodity is the translation layer that turns a reading into a change, and that layer finally has tooling. The download numbers on the structural side suggest people are noticing: the fastest-moving SEO skill in our directory isn't the one that generates content. It's the one that finds the redirect chain.