Performance Audit Skill
brunnocarpena/performance-audit-skill
Audit Lighthouse sem quebrar tracking, conversão ou layout
Ver no GitHubSkill para auditoria de desempenho de páginas usando Google PageSpeed Insights e Lighthouse, sem comprometer tracking, conversões ou layout. Fornece recomendações priorizadas por impacto real, custo de implementação e risco, preservando funcionalidades críticas de negócio.
Instalação
npx convez-ai brunnocarpena/performance-audit-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auditar páginas existentes (landing, WordPress, HTML, Next.js)
- Revisar código antes de deploy garantindo bom score desde o início
- Investigar regressões de Core Web Vitals ou quedas de conversão
- Otimizar performance em múltiplas plataformas de IA (Claude, Cursor, Copilot, Gemini, Cline)
Veredito

Skill prática e bem pensada que resolve o dilema real entre otimização de performance e preservação de tracking/conversão. Suporte amplo a plataformas de IA e documentação multilíngue demonstram maturidade; recomendado para times que precisam auditar antes de deploy.
README
A tradução para o português deste README ainda está na fila.
Performance Audit — Score 90+ on Lighthouse Without Breaking Tracking or Conversion
Page performance auditing through the lens of Google PageSpeed Insights / Lighthouse, without sacrificing layout, tracking, UTMs, SEO or conversion.
What it does
Performance audit that separates what Lighthouse measures from what the user actually feels, and prioritizes fixes with:
- Measurable impact on the right metric (LCP, CLS, INP, TBT, FCP).
- Realistic implementation cost.
- Explicit risk (breaks layout? breaks tracking? breaks conversion?).
- Post-fix test plan.
It is not blind optimization. Every recommendation preserves:
- Visual layout.
- Tracking (Pixel, GTM, GA4, Hotmart, AC, Make, conversions).
- UTMs and URL parameters.
- Form functionality.
- SEO (meta, structured data, canonical, hreflang).
- Basic accessibility.
Operating modes
audit— audit an existing page (URL, code, Lighthouse report).build— review code before deploy, ensuring performance from the start.
When to use
Use whenever:
- You need to audit a page (landing, WP/Elementor, HTML/Tailwind, Next.js, capture).
- You're building a new page and want to guarantee a good score from the start.
- Before any public-page deploy.
- Investigating a Core Web Vitals regression.
- Investigating a conversion drop suspected to be speed-related.
Invoke with /performance-audit or "audit this page's performance".
Installation
# 1. Clone into ~/.claude/skills/
mkdir -p ~/.claude/skills
cd ~/.claude/skills
git clone https://github.com/brunnocarpena/performance-audit-skill.git performance-audit
# 2. (Optional but recommended) configure PageSpeed Insights API key
# Without key: ~1 query/day keyed by IP. With key: 400 QPS / 25k queries/day.
#
# - https://console.cloud.google.com/apis/credentials → create API key
# - Enable "PageSpeed Insights API" and "Chrome UX Report API"
# - Add to ~/.zshrc (or ~/.bashrc):
#
# export PSI_API_KEY="AIza..."
# export CRUX_API_KEY="$PSI_API_KEY"
#
# - source ~/.zshrc
Validate installation:
ls ~/.claude/skills/performance-audit/SKILL.md # must exist
echo $PSI_API_KEY # should print the key (optional)
In any Claude Code session:
/performance-audit https://example.com --mode=audit --target=mobile
Supported AI platforms
| Platform | Adapter | Status |
|---|---|---|
| Claude Code (CLI / IDE / web) | root SKILL.md | ✅ Primary |
| Cursor | adapters/cursor/ | ✅ Supported |
| GitHub Copilot | adapters/copilot/ | ✅ Supported |
| Gemini CLI / Antigravity | adapters/gemini/ | ✅ Supported |
| OpenAI Codex CLI | adapters/codex/ | ✅ Supported |
| Aider | adapters/aider/ | ✅ Supported |
| Qwen Code CLI (Alibaba) | adapters/qwen/ | ✅ Supported |
| Kimi K2 (Moonshot AI) | adapters/kimi/ | ✅ Supported |
| Windsurf (Codeium) | adapters/windsurf/ | ✅ Supported |
| Cline (VS Code) | adapters/cline/ | ✅ Supported |
| Continue.dev | adapters/continue/ | ✅ Supported |
| Zed Editor | adapters/zed/ | ✅ Supported |
| JetBrains Junie | adapters/jetbrains-junie/ | ✅ Supported |
| Sourcegraph Cody | adapters/cody/ | ✅ Supported |
| ChatGPT Custom GPT | adapters/chatgpt-custom-gpt/ | ✅ Supported |
| Any other LLM | adapters/generic-prompt/ | ✅ Copy-paste |
Each adapter folder has its own INSTALL.md with platform-specific install steps.
When NOT to use
- SEO content audit →
seo-audit. - WCAG accessibility audit → dedicated skill.
- Security audit →
security-auditor. - Visual bug unrelated to performance.
Structure
performance-audit/
├── SKILL.md # Main instructions, Iron Law, workflow (EN canonical)
├── SKILL.pt-BR.md # Portuguese (Brazil)
├── SKILL.es.md # Spanish
├── SKILL.fr.md # French
├── SKILL.de.md # German
├── SKILL.zh.md # Simplified Chinese
├── README.md / README.*.md # This file in 6 languages
├── adapters/ # 15 platform adapters
├── references/ # Deep guides
│ ├── metrics-glossary.md # LCP/CLS/INP/TBT/FCP/SI/TTFB
│ ├── data-collection.md # How to collect lab + field data
│ ├── severity-rubric.md # Critical/high/medium/low rubric
│ ├── safe-fixes.md # Fixes by risk class
│ ├── report-format.md # Report template
│ └── psi-lighthouse-internals.md # PSI/Lighthouse/CrUX deep mechanics
└── checks/ # Domain checks
├── hero-and-lcp.md
├── images.md
├── fonts.md
├── css.md
├── javascript.md
├── third-party-tracking.md
├── wordpress-elementor.md
├── layout-stability.md
└── network-server.md
How it works
- Identifies scope and mode (audit vs build).
- Collects evidence: Lighthouse + PageSpeed + CrUX + code.
- Classifies into 12 categories (hero, images, fonts, CSS, JS, third-party, WP, CLS, network).
- Assigns affected metric + severity + priority + risk.
- Recommends safe fixes, always pointing to file/line and what NOT to touch alongside.
- Emits a standardized report.
- Decides pre-deploy: approved / with caveats / blocked.
Inviolable rules
- The LCP image never has
loading="lazy"— alwaysfetchpriority="high". - Tracking is not touched without explicit user confirmation.
- Lab and field data always separate — Lighthouse score is not absolute truth.
- Mobile and desktop always separate — they're different Google rankings.
- UTMs are never dropped in redirects/canonical.
- Every recommendation cites file/line or audit ID.
- Mobile-first by default.
- Report always emitted, even when everything is green.
- An isolated score does not approve deploy — Core Web Vitals "good" in field is the criterion.
- Every risky fix has a rollback plan.
Metrics and thresholds (2026)
| Metric | Good | Needs improvement | Poor |
|---|---|---|---|
| LCP | ≤ 2.5s | 2.5–4.0s | > 4.0s |
| CLS | ≤ 0.1 | 0.1–0.25 | > 0.25 |
| INP | ≤ 200ms | 200–500ms | > 500ms |
| TBT | ≤ 200ms | 200–600ms | > 600ms |
| FCP | ≤ 1.8s | 1.8–3.0s | > 3.0s |
| TTFB | ≤ 800ms | 800–1800ms | > 1800ms |
Google criterion: p75 ≥ "good" for 75% of users in CrUX.
Inspiration and references
Structure inspired by:
- brunnocarpena/congruence-skill — mandatory-evidence + Iron Law pattern.
- addyosmani/web-quality-skills — official Chrome team Agent Skills for Lighthouse + CWV.
- web.dev/articles/vitals — canonical source of thresholds.
- GoogleChrome/lighthouse-ci — budgets and regressions in PR.
Related skills
- congruence — verifies that page claims match the code. Use after if the audit modified copy/CTA.
- seo-audit — technical and on-page SEO. Complementary.
- ux-ui-perfectionist — UI/UX review. Use together if the optimization touches layout.
License
MIT — see LICENSE.