Website Care · Starter

Find what makes a page slow

Pages feel slow on phones, or Core Web Vitals in Search Console are failing.

  • Any
  • Claude Code
  • Grok Build
  • Cursor
  • Codex CLI

For: Owner, Office manager, Tech person, Anyone at work · Updated November 10, 2026

A starting point: fill in the [brackets] with your own details, and check the answer before you use it. Tool screens change often, so yours may look a little different.

The prompts

Copy a quick prompt, or grab the full skill below.

QUICK PROMPT 1
Why this prompt works: Asking for the top three forces prioritizing instead of a long list.
Measure how fast the home page of the site I work on loads on a phone. Name the three biggest things slowing it down, in plain language, before changing anything.
QUICK PROMPT 2
Why this prompt works: Give criteria: published targets make 'fast enough' measurable.
Using Google's Core Web Vitals targets (LCP under 2.5 s, CLS under 0.1, INP under 200 ms), check one article and the home page of the site I work on. Report pass or fail for each first.
QUICK PROMPT 3
Why this prompt works: A specific symptom gets a specific diagnosis.
Does anything on the site I work on jump around while loading? Tell me which elements shift and why, with no fixes yet.

What this covers: Measure and improve loading speed (LCP), responsiveness (INP), and layout shift (CLS): images, fonts, scripts, third parties, and caching, with before and after numbers.

For anyone improving a site or app, whether a small-business owner or someone at a larger company practicing better prompts. Slow pages lose visitors and rank lower in search, and the biggest wins are usually a few images, fonts, or scripts. Paste a prompt into any AI assistant that can see the site or repo.

Two ways to use it

  1. Any AI chat (Copilot, ChatGPT, Claude, Grok): copy one of the quick prompts above. Each one asks for a plain-language report first, so nothing changes until you say so.
  2. A coding agent working in the site’s code (Claude Code, Grok Build, Cursor, Codex CLI): save the full skill below as SKILL.md in a folder named performance-web-vitals-pass, for example .claude/skills/performance-web-vitals-pass/SKILL.md (Grok Build also reads that location, or use .grok/skills/). Then ask the agent to use the performance-web-vitals-pass skill. It audits first, makes the smallest safe changes in a review branch, and ends with a report.

Download SKILL.md · All Website Care skills

The full skill (SKILL.md)

Last reviewed Nov 10, 2026. Standards and tools change, so check anything dated before relying on it.

---
name: performance-web-vitals-pass
description: Use when asked to make a website or web app faster or to improve Core Web Vitals (LCP, INP, CLS) - images, fonts, JavaScript and CSS weight, third-party scripts, caching, and server response - using field data where available and lab data to diagnose. Delivers measured before/after numbers in a review PR. Not for visual redesign or feature work.
---

# Performance and Core Web Vitals pass

Last reviewed: 2026-11-10

## Quick prompts

For anyone improving a site or app, whether a small-business owner or someone at a larger company practicing better prompts. Slow pages lose visitors and rank lower in search, and the biggest wins are usually a few images, fonts, or scripts. Paste a prompt into any AI assistant that can see the site or repo.

1. "Measure how fast the home page of the site I work on loads on a phone. Name the three biggest things slowing it down, in plain language, before changing anything."
   *Why this prompt works:* Asking for the top three forces prioritizing instead of a long list.

2. "Using Google's Core Web Vitals targets (LCP under 2.5 s, CLS under 0.1, INP under 200 ms), check one article and the home page of the site I work on. Report pass or fail for each first."
   *Why this prompt works:* Give criteria: published targets make 'fast enough' measurable.

3. "Does anything on the site I work on jump around while loading? Tell me which elements shift and why, with no fixes yet."
   *Why this prompt works:* A specific symptom gets a specific diagnosis.

## Targets (75th percentile of real page loads, mobile and desktop separately)

| Metric | Good | Poor | Measures |
|---|---|---|---|
| LCP, Largest Contentful Paint | ≤ 2.5 s | > 4.0 s | Loading of the main content |
| INP, Interaction to Next Paint | ≤ 200 ms | > 500 ms | Responsiveness to taps, clicks, keys |
| CLS, Cumulative Layout Shift | ≤ 0.1 | > 0.25 | Visual stability |

Supporting: TTFB (server/CDN), FCP, and TBT (lab proxy for INP, since lab tools cannot measure INP without real input).

## Hard limits

- No visible design, content, or behavior changes. Lazy-loading or deferring must not hide content users need, break analytics the owner relies on, or remove a third-party tool without the owner's say.
- Review PR only; host/CDN cache rules, plan upgrades, or paid services need the owner's approval.
- Measure before and after with the same method; do not claim wins from a single noisy run.

## Inputs to gather

1. URL list: one per template, heaviest pages first (home, top landing pages, article, product, dashboard).
2. Field data if available: PageSpeed Insights (CrUX), Search Console Core Web Vitals report, or the site's own RUM (`web-vitals` library). Small sites often have no CrUX data; say so.
3. Stack, image pipeline, font loading, CDN, cache headers, third-party tags.
4. Build/deploy limits (one push per task if metered).

## Step 1: Classify

| Type | Likely bottlenecks | Careful with |
|---|---|---|
| Article / editorial | Hero image as LCP, web fonts, ads and embeds causing CLS, comment widgets | Ad slots and embeds the owner earns from: reserve space, do not remove |
| Marketing / landing | Hero video/image, sliders, tag managers, A/B testing scripts (render-blocking) | Tracking pixels: defer, do not delete without approval |
| Docs | Search index size, syntax highlighter JS, large sidebars | Search behavior |
| E-commerce | Product images, filters re-rendering lists (INP), third-party reviews/chat widgets | Checkout and payment scripts: do not defer or alter |
| Dashboard / SPA | JS bundle size, hydration cost, long tasks on data load (INP), chart libraries | Data freshness and polling intervals |
| SSR/static vs SPA | Static: images, fonts, caching. SPA: code splitting, route-level lazy loading, hydration | — |

## Step 2: Measure baseline

- Field: record p75 LCP/INP/CLS per device type if available.
- Lab: Lighthouse (mobile preset) 3–5 runs per URL, report the median; Chrome DevTools Performance panel for long tasks and the LCP element; WebPageTest for filmstrips if needed.
- Identify the LCP element on each template and what delays it (TTFB, resource load delay, load time, render delay).
- Record page weight by type (JS, CSS, images, fonts, third-party).

## Step 3: Fix (highest impact first)

**LCP**: make the LCP image discoverable in HTML (not CSS background or JS-injected), `fetchpriority="high"` on it, never `loading="lazy"` on it; responsive `srcset`/`sizes`; AVIF/WebP; preconnect to required origins; cut render-blocking CSS/JS; cache HTML at the CDN where safe.

**CLS**: `width`/`height` or `aspect-ratio` on images, videos, iframes and ad slots; reserve space for banners and embeds; `font-display: swap` or `optional` with size-adjusted fallbacks; avoid inserting content above existing content.

**INP**: break up long tasks (yield to main thread), defer non-critical JS, reduce hydration (islands, partial hydration), debounce heavy input handlers, keep DOM size reasonable, move work off the main thread.

**Weight and caching**: remove unused JS/CSS, code-split by route, subset and self-host fonts (WOFF2), lazy-load below-the-fold images and embeds (facades for video players), long `Cache-Control` with fingerprinted filenames, compression (Brotli/gzip).

**Third parties**: inventory each tag, load async/defer, delay non-essential ones until interaction or idle, and list candidates for removal for the owner to decide.

## Validation

- Re-run the same lab method on the same URLs; report medians before/after.
- Visual check in a real browser at desktop and phone sizes: nothing missing, no new layout shift, images sharp.
- Build and tests pass; analytics and forms still fire (in test mode).
- Field data takes about 28 days to reflect changes in CrUX; say when to recheck.

## Report format

1. Interface type(s), URLs, field data availability.
2. Table per URL: LCP element, LCP/CLS/TBT (and INP if field) before → after, page weight before → after.
3. Changes made, each with the metric it targets.
4. Third-party scripts and recommendations needing owner decision.
5. Not verified, and when to check field data. PR link, "not merged".

## References

- Web Vitals: https://web.dev/articles/vitals
- Optimize LCP / INP / CLS: https://web.dev/articles/optimize-lcp , https://web.dev/articles/optimize-inp , https://web.dev/articles/optimize-cls
- Most effective ways to improve Core Web Vitals: https://web.dev/articles/top-cwv
- Lab vs field data: https://web.dev/articles/lab-and-field-data-differences
- PageSpeed Insights: https://pagespeed.web.dev/
- MDN lazy loading and `fetchpriority`: https://developer.mozilla.org/en-US/docs/Web/Performance/Guides/Lazy_loading , https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Elements/img#fetchpriority
- MDN HTTP caching: https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/Caching

Check before you send

  • It gave a plain-language report before suggesting or making any change
  • Each finding names the exact page or file and why it matters
  • It said what it could not check or verify
  • Nothing was changed, published, or deleted without your OK
  • Fixes are small and keep the existing design and content

Privacy note: what not to paste

Paste page addresses, screenshots, or reports, not passwords, API keys, customer data, or private admin links. Only check sites you or your employer control, and get the site owner’s OK before anyone changes anything.