Generated banner for Performance
Generated logo for Performance

Performance

Addy OsmaniCommunityCreated by web-quality-skillsDistributed by Addy Osmani

  • Skill
  • Tool integration
  • Software development
  • MIT
  • No issues found

From the author

Optimize web performance for faster loading and better user experience. Use when asked to "speed up my site", "optimize performance", "reduce load time", "fix slow loading", "improve page speed", or "performance audit".
— the description by web-quality-skills

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a structured methodology for diagnosing and improving web performance. It emphasizes a data-driven approach, prioritizing real-user Core Web Vitals and using browser performance traces to identify specific bottlenecks. The workflow covers critical rendering path optimization, resource loading strategies, image and font optimization, caching, and runtime performance, ensuring that changes are based on measured evidence rather than assumptions.

What it does

  • Establishes performance baselines using field and lab data.
  • Identifies bottlenecks through browser performance traces and focused insights.
  • Provides optimization patterns for the critical rendering path and resource loading.
  • Offers guidance on image, font, and third-party script optimization.
  • Defines strategies for runtime performance, including layout and animation efficiency.
  • Outlines a measurement-first workflow for verifying improvements.

When to use it

  • When a site is experiencing slow loading times or poor user experience.
  • When you need to perform a performance audit on a web application.
  • Before and after implementing performance-related code changes to verify impact.
  • When you need to establish a performance budget for a project.

Advantages

  • Reduces guesswork by relying on real-user signals and diagnostic traces.
  • Provides actionable, evidence-led optimization patterns.
  • Helps maintain performance budgets across different device and network profiles.
  • Encourages a disciplined approach to measuring and verifying improvements.

Usage examples

“My website is loading slowly. Can you help me optimize it?”
The agent will initiate the performance measurement workflow, starting with a baseline assessment using field and lab data to identify specific bottlenecks.
“How can I improve the Largest Contentful Paint (LCP) for my landing page?”
The agent will analyze the critical rendering path, suggest preloading the LCP image, and check for render-blocking resources that might be delaying the paint.
“I need to reduce the total page weight of my application.”
The agent will review current resource usage against the recommended performance budget and suggest optimizations like code splitting, tree shaking, and image format conversion.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/addyosmani/performance/download
  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip performance.zip -d .agents/skills

Documentation by web-quality-skills

Performance optimization

Evidence-led performance optimization using real-user signals for prioritization and browser traces for diagnosis. Focuses on loading speed, runtime responsiveness, and resource delivery.

How it works

  1. If a page can run, read the measurement workflow and establish a field-plus-lab baseline before editing.

  2. Prioritize poor real-user Core Web Vitals. Use a DevTools performance trace and its focused insights to find the cause.

  3. Inspect and change only the code or assets connected to measured bottlenecks.

  4. Re-run equivalent lab measurements and report before/after values, conditions, and uncertainty. Field verification remains pending until enough new user data arrives.

When no runnable page exists, perform static inspection but call findings hypotheses, not measured regressions. Include the command or browser workflow that can verify each high-impact hypothesis.

Prefer a browser tool that records a performance trace and exposes focused insights. With Chrome DevTools MCP, use performance_start_trace and performance_analyze_insight; do not route performance through lighthouse_audit, which covers non-performance Lighthouse categories.

Starting performance budget

Budgets must reflect the product's target devices, networks, page types, and user journeys. The values below are initial guardrails for a typical content or commerce page, not universal pass/fail criteria. Preserve an existing project budget when one is already defined.

ResourceBudgetRationale
Total page weight< 1.5 MBBounds transfer time and data cost on constrained target networks; calibrate with representative pages
JavaScript (compressed)< 300 KBProtect parse and execution cost
CSS (compressed)< 100 KBLimit render-blocking work
Images (above-fold)< 500 KBProtect likely LCP resources
Fonts< 100 KBLimit critical font transfer
Third-party< 200 KBBound code outside product control

Critical rendering path

Server response
  • TTFB < 800ms. Time to First Byte should be fast. Use CDN, caching, and efficient backends.

  • Enable compression. Gzip or Brotli for text assets. Brotli preferred (15-20% smaller).

  • HTTP/2 or HTTP/3. Multiplexing reduces connection overhead.

  • Edge caching. Cache HTML at CDN edge when possible.

  • Consider Early Hints (HTTP 103) for measured document latency. If a trace shows slow HTML generation and stable critical subresources, send an interim 103 with Link headers before the normal final response from the same request. Use HTTP/2 or later. A CDN may synthesize the 103 from Link headers on an earlier 200, or the origin/edge handler can emit it directly. Unsupported clients continue to the final response, but confirm current browser and infrastructure support. Limit hints to proven critical preloads or preconnects: inaccurate hints waste bandwidth. Cloudflare reported a 20–30% LCP improvement in an artificial, image-heavy test; treat that as a vendor case study, not an expected saving, and measure your result. See MDN's 103 implementation example and the Cloudflare study.

Resource loading

Preconnect to required origins:

<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://cdn.example.com" crossorigin>

Preload critical resources:

Preload only resources whose late discovery is visible in the trace. Each preload competes for bandwidth and an unnecessary high-priority request can delay LCP.

<!-- LCP image -->
<link rel="preload" href="/hero.webp" as="image" fetchpriority="high">

<!-- Critical font -->
<link rel="preload" href="/font.woff2" as="font" type="font/woff2" crossorigin>

Prerender likely-next navigations with the Speculation Rules API:

<script type="speculationrules">
{
  "prerender": [{
    "where": { "href_matches": "/*" },
    "eagerness": "moderate"
  }]
}
</script>

moderate waits for a stronger intent signal than eager modes. Measure prediction hit rate, transferred bytes, and server cost; a wrong prerender is roughly an unused navigation. See core-web-vitals → LCP for the tradeoffs and the prerenderingchange gating needed for analytics.

Defer non-critical CSS:

<!-- Critical CSS inlined -->
<style>/* Above-fold styles */</style>

<!-- Non-critical CSS -->
<link rel="preload" href="/styles.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<noscript><link rel="stylesheet" href="/styles.css"></noscript>
JavaScript optimization

Defer non-essential scripts:

<!-- Parser-blocking (avoid) -->
<script src="/critical.js"></script>

<!-- Deferred (preferred) -->
<script defer src="/app.js"></script>

<!-- Async (for independent scripts) -->
<script async src="/analytics.js"></script>

<!-- Module (deferred by default) -->
<script type="module" src="/app.mjs"></script>

Code splitting patterns:

// Route-based splitting
const Dashboard = lazy(() => import('./Dashboard'));

// Component-based splitting
const HeavyChart = lazy(() => import('./HeavyChart'));

// Feature-based splitting
if (user.isPremium) {
  const PremiumFeatures = await import('./PremiumFeatures');
}

Tree shaking best practices:

// ❌ Imports entire library
import _ from 'lodash';
_.debounce(fn, 300);

// ✅ Imports only what's needed
import debounce from 'lodash/debounce';
debounce(fn, 300);

Image optimization

Format selection
FormatUse caseBrowser support
AVIFPhotos, best compression92%+
WebPPhotos, good fallback97%+
PNGGraphics with transparencyUniversal
SVGIcons, logos, illustrationsUniversal
Responsive images
<picture>
  <!-- AVIF for modern browsers -->
  <source 
    type="image/avif"
    srcset="hero-400.avif 400w,
            hero-800.avif 800w,
            hero-1200.avif 1200w"
    sizes="(max-width: 600px) 100vw, 50vw">
  
  <!-- WebP fallback -->
  <source 
    type="image/webp"
    srcset="hero-400.webp 400w,
            hero-800.webp 800w,
            hero-1200.webp 1200w"
    sizes="(max-width: 600px) 100vw, 50vw">
  
  <!-- JPEG fallback -->
  <img 
    src="hero-800.jpg"
    srcset="hero-400.jpg 400w,
            hero-800.jpg 800w,
            hero-1200.jpg 1200w"
    sizes="(max-width: 600px) 100vw, 50vw"
    width="1200" 
    height="600"
    alt="Hero image"
    loading="lazy"
    decoding="async">
</picture>
LCP image priority
<!-- Above-fold LCP image: eager loading, high priority -->
<img 
  src="hero.webp" 
  fetchpriority="high"
  loading="eager"
  decoding="sync"
  alt="Hero">

<!-- Below-fold images: lazy loading -->
<img 
  src="product.webp" 
  loading="lazy"
  decoding="async"
  alt="Product">

Font optimization

Loading strategy
/* System font stack as fallback */
body {
  font-family: 'Custom Font', -apple-system, BlinkMacSystemFont, 
               'Segoe UI', Roboto, sans-serif;
}

/* Prevent invisible text */
@font-face {
  font-family: 'Custom Font';
  src: url('/fonts/custom.woff2') format('woff2');
  font-display: swap; /* or optional for non-critical */
  font-weight: 400;
  font-style: normal;
  unicode-range: U+0000-00FF; /* Subset to Latin */
}
Preloading critical fonts
<link rel="preload" href="/fonts/heading.woff2" as="font" type="font/woff2" crossorigin>
Variable fonts
/* One file instead of multiple weights */
@font-face {
  font-family: 'Inter';
  src: url('/fonts/Inter-Variable.woff2') format('woff2-variations');
  font-weight: 100 900;
  font-display: swap;
}

Caching strategy

Cache-Control headers
# HTML (short or no cache)
Cache-Control: no-cache, must-revalidate

# Static assets with hash (immutable)
Cache-Control: public, max-age=31536000, immutable

# Static assets without hash
Cache-Control: public, max-age=86400, stale-while-revalidate=604800

# API responses
Cache-Control: private, max-age=0, must-revalidate
Service worker caching
// Cache-first for static assets
self.addEventListener('fetch', (event) => {
  if (event.request.destination === 'image' ||
      event.request.destination === 'style' ||
      event.request.destination === 'script') {
    event.respondWith(
      caches.match(event.request).then((cached) => {
        return cached || fetch(event.request).then((response) => {
          const clone = response.clone();
          caches.open('static-v1').then((cache) => cache.put(event.request, clone));
          return response;
        });
      })
    );
  }
});

Runtime performance

Avoid layout thrashing
// ❌ Forces multiple reflows
elements.forEach(el => {
  const height = el.offsetHeight; // Read
  el.style.height = height + 10 + 'px'; // Write
});

// ✅ Batch reads, then batch writes
const heights = elements.map(el => el.offsetHeight); // All reads
elements.forEach((el, i) => {
  el.style.height = heights[i] + 10 + 'px'; // All writes
});
Debounce expensive operations
function debounce(fn, delay) {
  let timeout;
  return (...args) => {
    clearTimeout(timeout);
    timeout = setTimeout(() => fn(...args), delay);
  };
}

// Debounce scroll/resize handlers
window.addEventListener('scroll', debounce(handleScroll, 100));
Use requestAnimationFrame
// ❌ May cause jank
setInterval(animate, 16);

// ✅ Synced with display refresh
function animate() {
  // Animation logic
  requestAnimationFrame(animate);
}
requestAnimationFrame(animate);
Virtualize long lists
// For lists > 100 items, render only visible items
// Use libraries like react-window, vue-virtual-scroller, or native CSS:
.virtual-list {
  content-visibility: auto;
  contain-intrinsic-size: 0 50px; /* Estimated item height */
}
Smooth navigations with View Transitions

The View Transitions API lets the browser cross-fade (or custom-animate) between two DOM states using a single GPU-composited snapshot — no double-render, no layout thrash, and the snapshot doesn't count toward CLS.

Same-document (SPA-style) — Baseline 2026:

// Wrap the DOM mutation that swaps the view
function navigate(newView) {
  if (!document.startViewTransition) return swapDOM(newView);
  document.startViewTransition(() => swapDOM(newView));
}

Cross-document (MPA-style) — Chromium-stable, progressive enhancement elsewhere:

/* On both source and destination pages */
@view-transition { navigation: auto; }

That's the entire integration — same-origin navigations now fade automatically. To opt specific elements into shared-element transitions (e.g. a thumbnail expanding into a hero), give them a matching view-transition-name:

.product-thumb[data-id="42"], .product-hero { view-transition-name: product-42; }

Pair this with Speculation Rules (above) for instant + animated navigations.

Third-party scripts

Load strategies
// ❌ Blocks main thread
<script src="https://analytics.example.com/script.js"></script>

// ✅ Async loading
<script async src="https://analytics.example.com/script.js"></script>

// ✅ Delay until interaction
<script>
document.addEventListener('DOMContentLoaded', () => {
  const observer = new IntersectionObserver((entries) => {
    if (entries[0].isIntersecting) {
      const script = document.createElement('script');
      script.src = 'https://widget.example.com/embed.js';
      document.body.appendChild(script);
      observer.disconnect();
    }
  });
  observer.observe(document.querySelector('#widget-container'));
});
</script>
Facade pattern
<!-- Show static placeholder until interaction -->
<div class="youtube-facade" 
     data-video-id="abc123" 
     onclick="loadYouTube(this)">
  <img src="/thumbnails/abc123.jpg" alt="Video title">
  <button aria-label="Play video">▶</button>
</div>

Measurement

Use the measurement workflow whenever a URL is runnable. It defines Chrome DevTools MCP routing, CrUX and fallback sources, repeatable lab conditions, and a compact evidence format.

MetricKindInterpretation
LCP, INP, CLS at p75FieldUser-outcome Core Web Vitals; use for pass/fail prioritization
LCP, CLS in a traceLabReproducible diagnostic values for one navigation
TBTLabMain-thread blocking diagnostic and a rough INP proxy, not field INP
FCP, Speed IndexLabLoading diagnostics, not Core Web Vitals

Raw PerformanceObserver snippets are useful for the current browser session but are not real-user data by themselves. When the user wants production telemetry, read the first-party RUM reference and prefer web-vitals over a hand-rolled metric implementation.

References

For Core Web Vitals specific optimizations, see Core Web Vitals.

Files

references/MEASUREMENT.md

Performance measurement workflow

Use this workflow when a runnable URL is available, the user asks for measured performance, or a change needs before/after verification.

Keep the evidence types separate

EvidenceWhat it representsBest use
CrUX field dataAggregated experiences from eligible real Chrome users, normally a rolling 28-day windowDecide whether users have a Core Web Vitals problem
First-party RUMMeasurements collected and reported from the site's own user sessionsSegment and diagnose current production experience
DevTools performance traceOne observed browser session under stated local or emulated conditionsFind LCP, INP, CLS, network, and main-thread causes
Lighthouse lab runA controlled synthetic navigationReproduce load problems and prevent regressions
Static code inspectionPotential issues inferred from sourceForm hypotheses when no page can run

A PerformanceObserver result injected into one browser page is a single-session lab observation, not field data. It becomes RUM only when measurements from actual users are reported and aggregated.

Preferred low-friction route

When browser tooling can record a performance trace and run Lighthouse audits, prefer this route. With Chrome DevTools MCP:

  1. Navigate to the exact route and state being audited. Record whether it is public, authenticated, local, or staging.

  2. Record a reload trace with auto-stop for page-load performance (performance_start_trace). Current trace summaries can include both observed lab metrics and CrUX field metrics when CrUX has eligible data. Record whether field scope is URL or origin.

  3. Analyze only the relevant failing or suspicious insights (performance_analyze_insight). Common examples are LCPBreakdown, LCPDiscovery, DocumentLatency, RenderBlocking, and ThirdParties.

  4. Run the Lighthouse audit capability (lighthouse_audit) for Accessibility, SEO, Best Practices, and Agentic Browsing. It deliberately excludes performance; do not treat it as the performance path.

  5. Re-run the same lab measurement after a fix. Field data will not reflect a new deployment immediately.

Use mobile conditions by default for a general public-site audit. Add desktop when the user asks for it, desktop traffic matters, or the product is desktop-oriented. Test authenticated and unauthenticated states separately when they render different pages.

When using emulation, set the viewport, network conditions, and CPU rate explicitly before the trace, confirm the reported conditions, and reset them before testing another profile.

Chrome DevTools MCP's Lighthouse navigation mode reloads the page. Use snapshot mode for the current state when a reload would lose an authenticated or user-created state. Do not run a navigation audit on an unsaved form or destructive workflow.

Token-efficient tool use
  • Start with one trace and one Lighthouse audit rather than broad DOM, network, console, and source dumps.

  • Preserve large reports or traces to temporary files when the tool supports filePath or outputDirPath; summarize only actionable failures.

  • Drill into the few insights tied to a poor field metric or a reproducible lab bottleneck.

  • Filter and paginate network or console requests. Fetch individual request details only when they support a finding.

  • Take a text snapshot before a screenshot unless visual inspection is necessary.

Fallbacks when DevTools tools are unavailable

Use the first available option; do not block the audit on optional setup.

  1. Lighthouse CLI for lab data: run the project's compatible Lighthouse version against the runnable URL. Keep JSON for comparison and avoid installing a permanent dependency unless the user wants one.

  2. PageSpeed Insights web UI for a public URL: it provides a zero-setup view of Lighthouse lab diagnostics and available CrUX field data.

  3. CrUX Vis for history: use it when trend data matters and the URL or origin is eligible.

  4. CrUX API or History API for automation: both are free to use but require a Google Cloud API key. Do not make a key a prerequisite for an ordinary audit.

  5. Static inspection: if nothing can run, label every performance finding as a hypothesis and provide the exact measurement needed to verify it.

The PageSpeed Insights API may be called without a key for occasional use, but a key is recommended for repeated automation. Google has announced that CrUX field data will be removed from that API, so new integrations should query the CrUX API directly.

Reading CrUX correctly

  • Prefer page-level data for the audited URL. If only origin data exists, label it as origin scope; it is context, not proof for that route.

  • Compare the p75 value with the Core Web Vitals threshold and include the percentage of good experiences when available.

  • Keep phone and desktop data separate. Do not combine form factors to answer a device-specific question.

  • Treat missing CrUX data as unavailable, never as passing. Localhost, staging, new, private, and low-traffic pages commonly have no CrUX record.

  • CrUX is aggregated and delayed. Use it to prioritize user outcomes, not to verify a change deployed minutes ago.

Repeatable lab comparisons

Record these conditions with the result:

  • final URL and page state

  • browser and Lighthouse/tool version

  • viewport or form factor

  • CPU and network throttling

  • cold or warm cache

  • authentication, consent, and experiment state

For a decision based on a headline lab metric, run at least three equivalent navigations and report the median plus range. Do not compare a single local trace directly with the CrUX p75 or claim that the two should match.

Use metric values as the evidence. A Lighthouse score is a diagnostic summary whose weighting and implementation can change between versions.

Reconciling lab and field

FieldLabInterpretation
PoorPoorReproducible user problem; trace and fix the dominant bottleneck
PoorGoodLocal run missed real-user conditions; segment first-party RUM or test representative devices, routes, cache states, and interactions
GoodPoorThe synthetic cold/throttled case is fragile, but do not claim users are currently failing
UnavailableAnyUse lab data for diagnosis and recommend RUM if production impact matters

Compact audit output

Start reports with an evidence table:

SignalScope and conditionsBaselineAfterSource
LCPURL, phone, p75/28 days3.1sPending field windowCrUX
LCPURL, mobile lab, cold cache, median of 33.8s2.6sDevTools trace

Then separate:

  1. measured failures

  2. trace-backed causes

  3. source-code hypotheses

  4. fixes made or recommended

  5. verification status and remaining uncertainty

Sources

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Security

No issues found by our static scan· No issues found

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by
web-quality-skills
Publisher
Addy Osmani (Community)
Distributed by
Addy Osmani
Source
addyosmani/web-quality-skills · skills/performance · commit afa8da9
Licence
MIT License (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by web-quality-skills or Addy Osmani.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download performance.zip