Turn field data, lab reports, traces, and source evidence into prioritized web performance fixes with measurable release gates.
A Lighthouse score can move because the test machine changed, while real users still wait on the same slow route. Field data can flag a bad experience without showing which request, task, or component caused it. This skill keeps those evidence types separate and makes them work together.
It maps performance by route and user journey, reconciles field and lab results, traces each material symptom to an owned resource or code path, and ranks fixes by impact, confidence, effort, and regression risk. The output includes a release verdict, causal findings, a fix portfolio, owned budgets, and a verification matrix that protects behavior, accessibility, SEO, analytics, and cache correctness.
Use it for Core Web Vitals work, a slow page or interaction, a performance regression, or a release gate. A bundled Python tool extracts stable evidence from Lighthouse and PageSpeed Insights JSON without fetching a URL or changing the report. The package does not promise ranking or conversion gains, treat TBT as measured INP, or call a fix verified while field data is still pending.
Input
PageSpeed Insights shows mobile field LCP at 4.2 seconds for the product route over the last 28 days, but one Lighthouse run on the homepage reports LCP at 1.9 seconds and a score of 92. The release changed product images and CDN rules. Tell us what to fix before Friday.
Output
A measure-first verdict; separate page, origin, route, field-window, and lab evidence; a product-route measurement contract; hypotheses for cache, geography, image discovery, and route coverage; discriminating tests; and a fix and verification matrix that does not erase poor field data with the unrelated homepage run.
Input
This checkout release raises Lighthouse from 78 to 86 after delaying the tag manager, lazy-loading the hero, and caching HTML at the CDN. We have one run before and one after. Checkout is personalized, fraud and consent tags share the manager, and the hero is usually LCP. Can we ship?
Provide the target routes and user journey, release or incident context, any CrUX or RUM field evidence, Lighthouse, PageSpeed Insights, trace, or bundle reports, and the relevant source or delivery configuration. Include measurement profiles, dates, revisions, budgets, and constraints when known. The skill returns a scoped verdict, separate field and lab baselines, causal findings, prioritized fixes, owned gates, and a comparable verification plan without inventing measurements or applying changes by default.
The origin, route or route cohort, device and geography context, and the user action that matters. A single URL can support a page-level audit when broader coverage is unavailable.
Field data, lab reports, traces, source, headers, build output, or monitoring records. Missing evidence stays explicit and can lead to a measure-first verdict.
The relevant build or release, existing budgets, supported devices, business-critical behavior, accessibility and SEO constraints, owners, and rollback options when known.
Required only for the bundled read-only JSON summarizer. It uses the standard library, writes to stdout, and does not require network access.
SKILL.md; field-versus-lab audit workflow; severity, evidence, causal-chain, budget, fix-quality, and verification rubric; read-only Lighthouse and PageSpeed Insights JSON summarizer; Agent Skills interface metadata
No reviews yet.
Output
A do-not-release or ready-with-conditions verdict tied to incomparable single runs, public-cache risk, LCP lazy-loading, and the unreviewed consent and fraud dependency; a split change plan; functional guardrails; repeated comparable lab checks; staged rollout; field-pending status; stop conditions; and rollback.
Input
RUM shows p75 INP at 340 ms on low-end Android for search filters. A trace captures a 420 ms click with 90 ms input delay, a 250 ms JavaScript task in filters.bundle.js, and 80 ms presentation delay. The bundle includes app filters and a third-party recommendations widget. Lighthouse TBT is 110 ms. Prioritize the work without promising a percentage improvement.
Output
A high-severity, evidence-bound responsiveness finding; the field INP and lab TBT kept separate; a causal chain from the filter interaction through the long task and render delay to named owners; contained experiments for app and third-party work; accessibility and analytics guardrails; and comparable interaction, field-cohort, rollout, stop, and rollback gates.
Creator
Mmarvex4
