# Lighthouse > Lighthouse audit of one page with an improvement loop until targets are met. Triggers "lighthouse", "page speed", "LCP", "CLS", "INP", "core web vitals". Repo-wide performance audits go to /audit performance. - Skill: `darkroomengineering/lighthouse` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add darkroomengineering/lighthouse` - Raw SKILL.md: https://api.skillmd.com/api/skills/darkroomengineering/lighthouse/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security - Author: darkroomengineering (https://skillmd.com/u/darkroomengineering) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/darkroomengineering/lighthouse --- # Lighthouse Optimization Loop Set a product-neutral scratch root once: ```bash LIGHTHOUSE_DIR="${TMPDIR:-/tmp}/cc-settings-lighthouse" mkdir -p "$LIGHTHOUSE_DIR" ``` Use the Chrome DevTools MCP only when the user configured it. Standalone Codex otherwise uses the Lighthouse CLI plus native/manual screenshots; if no visual capture path exists, stop before changing UI and report that visual regression verification is unavailable. This package does not auto-run unpinned registry MCP packages. **Method:** 3 mobile + 3 desktop runs per audit, averaged for reliability. After each code change, re-audit and visually verify the page with the configured MCP or the stated manual fallback. --- ## Setup 1. **Parse URL** from `$ARGUMENTS`. If no URL, ask the user. Default: `http://localhost:3000` 2. **Verify prerequisites:** ```bash lighthouse --version # CLI must be installed for the batched 3x3 protocol ``` If lighthouse is missing: `npm install -g lighthouse` If runs fail with a chrome-launcher error or every metric comes back `NO_LCP`/null, no system Chrome exists — find a real Chrome/Chromium binary (Chrome for Testing, Playwright's `chromium-*` — NOT `chrome-headless-shell`, which produces NO_LCP) and export `CHROME_PATH=` before the loop. Record which binary was used; a before/after comparison is only clean when both sides ran the same binary. Check whether the user configured the chrome-devtools MCP. Do not install or auto-run an unpinned registry MCP on their behalf. 3. **Create results directory:** ```bash mkdir -p "$LIGHTHOUSE_DIR" ``` 4. **Take baseline screenshots** before any changes: - `mcp__chrome-devtools__navigate_page` (type: "url", url: ``) - `mcp__chrome-devtools__take_screenshot` Describe the current layout, key elements, and visual state. This is your **visual baseline** — you will compare against it after every change to catch regressions. 5. **Confirm with user:** Show the URL, confirm the dev server is running, ask if there are specific pages or routes to audit beyond the main URL. --- ## Audit Protocol Each audit consists of **3 mobile + 3 desktop runs**, averaged per category. ### Run the audits ```bash # Mobile runs (Lighthouse default is mobile) for i in 1 2 3; do lighthouse \ --output=json \ --output-path="$LIGHTHOUSE_DIR/mobile-$i.json" \ --chrome-flags="--headless --no-sandbox" \ --only-categories=performance,accessibility,best-practices,seo \ --quiet \ 2>/dev/null done # Desktop runs for i in 1 2 3; do lighthouse \ --output=json \ --output-path="$LIGHTHOUSE_DIR/desktop-$i.json" \ --chrome-flags="--headless --no-sandbox" \ --preset=desktop \ --only-categories=performance,accessibility,best-practices,seo \ --quiet \ 2>/dev/null done ``` ### Extract scores For each JSON result file: ```bash cat "$LIGHTHOUSE_DIR/mobile-1.json" | \ jq '{ performance: (.categories.performance.score * 100), accessibility: (.categories.accessibility.score * 100), bestPractices: (.categories["best-practices"].score * 100), seo: (.categories.seo.score * 100) }' ``` ### Compute averages Average the 3 runs per category for both mobile and desktop. Report as: ``` ## Audit Results | Category | Mobile (avg) | Desktop (avg) | |----------|-------------|---------------| | Performance | XX | XX | | Accessibility | XX | XX | | Best Practices | XX | XX | | SEO | XX | XX | ``` ### Extract failing audits From the JSON, find specific audits that failed or scored poorly: ```bash cat "$LIGHTHOUSE_DIR/mobile-1.json" | \ jq '.audits | to_entries[] | select(.value.score != null and .value.score < 0.9) | {id: .key, score: .value.score, title: .value.title, description: .value.displayValue}' ``` Sort by impact (lowest scores first). These drive the improvement loop. --- ## Diagnose From the JSON, Not From a Narrative Before planning any fix, re-derive the bottleneck from the report you just ran. Prior session notes, state files, and issue write-ups go stale and get root causes wrong — treat them as hypotheses to check against the JSON, never as the diagnosis. **Classify the perf gap first:** - **TBT high (hundreds of ms)** → main-thread problem. Long tasks, hydration, script execution. `mainthread-work-breakdown` and `long-tasks` tell you where. - **TBT low (~tens of ms) but LCP high under `throttlingMethod=simulate`** → bytes-on-critical-path problem, NOT execution. Lantern replays observed traffic over simulated slow 4G, so eager bytes inflate LCP even when the page is actually fast. Fixes that defer *execution* will not move this; only removing or lazy-loading *bytes* will. Do not read `mainthread-work-breakdown`'s total as "blocking work before paint" — it is the whole page load. TBT is the blocking measure. **For bytes-on-critical-path gaps, attribute before fixing:** ```bash jq '.audits["resource-summary"].details.items[] | {resourceType, requestCount, transferSize}' "$LIGHTHOUSE_DIR/mobile-1.json" jq '.audits["bootup-time"].details.items[:5]' "$LIGHTHOUSE_DIR/mobile-1.json" jq '.audits["network-requests"].details.items | sort_by(-.transferSize)[:10] | .[] | {transferSize, resourceType, url}' "$LIGHTHOUSE_DIR/mobile-1.json" ``` For a full attribution of a route's JS by package and app directory, and the measured fix ladder for bytes-on-critical-path gaps (AVIF, per-route CSS chunking, font subsetting with `unicode-range`, platform elements over client widgets), use `~/.claude/skills/audit/references/performance-playbook.md` and its two harness scripts. If a hosted runner (PageSpeed Insights) reports "the page stopped responding" while local runs pass, the page runs WebGL on a software renderer: reproduce with `--chrome-flags='--headless=new --disable-gpu --use-gl=angle --use-angle=swiftshader'` and gate the canvas on the renderer string (playbook rung 7). Hunt for bytes that ship but do nothing: preloaded fonts no style consumes (a grep hit on the font file may be its own declaration — check for a real consumer), and libraries a `next/dynamic` wrapper claims to defer while a second static import path pulls them in anyway. Verify eager-path claims against the build manifest (`.next/build-manifest.json` on Next.js), not against the wrapper's existence. A `next/dynamic({ssr:false})` component that renders unconditionally still fetches its chunk at hydration — deferral only pays when the render is conditional. --- ## Improvement Loop > **Autonomous mode:** to drive this loop turn-by-turn without re-prompting, set > `/goal mobile and desktop scores in all four categories meet their targets, or stop after 20 rounds`. > A goal evaluator (Haiku by default) reads the audit table after each turn and decides whether to continue. > See [/goal docs](https://code.claude.com/docs/en/goal). ``` LOOP until all scores >= 90 or user interrupts: 1. IDENTIFY the lowest-scoring category and its top failing audits - Read the Lighthouse audit details for specific recommendations - Cross-reference with the project's performance rules 2. PLAN one targeted fix - Focus on the highest-impact failing audit - One fix at a time — never batch multiple unrelated changes - Common fixes by audit: • render-blocking-resources → async/defer scripts, inline critical CSS • largest-contentful-paint → priority attribute, preload, optimize image • cumulative-layout-shift → explicit dimensions, font-display • unused-javascript → dynamic imports, code splitting • uses-responsive-images → srcSet + sizes, `next/image` (satus) or `