Eight Agent Skills for Claude · Free & open source

Get the audit a consultant charges for.

Eight skills read your live site — messaging, conversion, SEO, UX, design — and hand back a graded, prioritised audit. Measured, not guessed.

Get the suite Read the source ↗
Example finding CRO-03
Issue
The homepage hero opens with “Welcome to our website.” — no offer, no CTA above the fold. The only button reads “Learn more.”
Impact
Visitors can’t tell what you sell or what to do next in the first five seconds — the highest-traffic moment on the site. Impact 5 / 5.
Fix
Replace with the outcome you deliver + one primary action, e.g. “Cut invoice processing to under a minute — Start free.” A one-line copy change. Effort 1 / 5.
Quick Win Messaging & Clarity graded D · 58/100 · rolled into the Action Report’s do-first roadmap.
01

Eight skills, one audit

Each skill owns one dimension and grades it A–F against a transparent rubric. Start with Snapshot; run any of the middle five; the Action Report ties them into a single plan.

0

Site Snapshot

Fetches the URL and captures what the site is, who it’s for, its primary conversion goal, and its key pages — the reusable foundation everything else builds on.

start here
1

Goals & Discovery

A short owner interview: are you happy with the site, what’s the one goal + a success number, what are the constraints. Runs while the scans render.

needs 0 · runs in parallel
2

Messaging & Clarity

Is the value proposition, headline, and call to action instantly clear? Graded, with MSG- findings and cross-page checks.

needs 0
3

Conversion (CRO)

Funnel, CTAs, forms, trust signals, distractions, pricing clarity — plus measured consent & conversion-tracking checks. CRO- findings.

needs 0
4

SEO & Content

Titles, meta, headings, keyword intent, content depth, schema, indexability — with an optional site-wide sweep. SEO- findings.

needs 0
5

UX & Technical

Navigation, measured mobile fit & tap-targets, real Core Web Vitals, accessibility, links, security — via a headless-browser render. UX- findings.

needs 0
6

Design & Visual

Renders desktop + mobile and judges hierarchy, type, colour, whitespace, modern-vs-dated — plus measured design tokens. DSN- findings.

needs 0
7

Action Report

The overall scorecard, an Impact × Effort priority matrix, a phased roadmap, and the single consolidated Website Audit Report — one file to act from.

needs 2–6

One URL in, one plan out.

Run the whole suite or just the dimension you care about. Nothing is invented — what can’t be measured is labelled, never faked.

The run order

Snapshot feeds everything. The interview runs alongside the scans, so the owner is being talked to while the machine works. The five analyses are independent; the Action Report synthesises whatever you ran.

0
Site Snapshotfetch · what / who / goal / key pages
│
1
Goals & Discoveryowner interview — runs in parallel with the scans below
▼  fans out to the five analyses
Run any / all — independent
02Messaging
03Conversion
04SEO
05UX / Tech
06Design
▼
7
Action Reportscorecard · Impact × Effort matrix · roadmap · consolidated report
Basic works anywhere Claude can browse — nothing to install. Extended adds a headless browser for the measured checks (Core Web Vitals, WCAG, mobile fit, design tokens). Both free.
02

Every finding is Issue · Impact · Fix

One repeatable unit underpins the whole suite — so nothing is vague opinion and everything is actionable.

I

Issue

The specific problem on the page, with the actual copy or element quoted as evidence. No hand-waving — you can see exactly what it points at.

I

Impact

What it costs the business — lost conversions, an unclear message, missed search traffic, friction — rated 1–5.

F

Fix

The concrete change to make, ideally with a worked example, plus an Effort 1–5. Impact × Effort places it on the priority board.

Impact × Effort tells you what to do first. Every finding lands in one quadrant, so the report isn’t a wall of problems — it’s an order of operations.
Low effort
High effort
High impact
Quick Windo first
Big Betplan & schedule
Low impact
Fill-inwhen convenient
MSG- messaging CRO- conversion SEO- search UX- ux/tech DSN- design
03

Two layers, both free

The suite runs in two layers, and both produce a real, graded audit. The only difference is a one-time local setup that unlocks measured evidence — hard numbers from a real render. You never lose anything by staying on Basic; Extended just fills those numbers in.

Basic — nothing to install

Claude reads & reasons

Runs in the Claude web app, desktop, or Claude Code with zero setup. Claude fetches your live site and reasons about it, producing the full graded audit: every finding, the priorities, the A–F scorecard, the 2026 comparison, and the styled, share-ready report. This is the expert-review layer — and on its own it already covers most of the audit.

  • → Works anywhere Claude can browse a URL
  • → Extended-only checks are honestly labelled “not measured”
Extended — adds hard evidence

A real browser renders & times the page

A one-time setup adds the checks Claude can’t do by only reading a page, because they need the page actually rendered and timed. These run through bundled tools and get filled straight into the same report. Each tool-backed finding carries a small measured tag, so you see exactly which evidence came from a real render.

  • → Real Core Web Vitals · full axe WCAG scan
  • → Measured mobile fit, 44px tap-targets, design tokens

Why Playwright (and its Chromium)?

The measured checks need your page actually rendered — laid out, styled, and timed by a real browser, not just read as HTML. So the suite drives one, and we chose Playwright deliberately: it’s the render engine that behaves identically on macOS, Linux, and Windows, and it downloads its own private copy of Chromium in a single command. One setup, same result on every OS — no system browser to install, nothing to configure, no API keys. Reading the HTML can tell you a <title> is missing; only a real render can tell you the page doesn’t fit a phone, takes four seconds to become usable, or fails a colour-contrast check. And if Playwright isn’t installed, every measured check simply falls back to not measured and the audit still runs — it never guesses, and never sends you to a third-party service.

DimensionBasic — no installExtended adds
Site SnapshotFull—
Goals & DiscoveryFull owner interview— (runs in parallel with the scans)
Messaging & ClarityFull, incl. cross-pageOptional screenshot-based read
Conversion (CRO)Full findings + CTA / link check + “is analytics installed?”Consent-before-load & cookie behaviour, runtime conversion-tracking proof
SEO & ContentFull — Claude reads the key pagesAutomated site-wide sweep — faster, more pages, exact
UX & TechnicalViewport / HTTPS, markup a11y, navigation, speed risk flagsReal Core Web Vitals · full axe accessibility scan · measured mobile fit & 44px tap-targets
Design & VisualDeclared colours / fonts; visual read if you paste a screenshotMeasured design tokens, auto desktop + mobile screenshots, cross-page consistency
Action ReportFull synthesis + 2026 scorecardSame, with the measured findings filled in
What the Extended tools measure

Static HTML can’t tell you whether a page fits a phone, how fast it is, whether it’s accessible, or what its design system looks like. Those need a real render — so the suite drives headless Chromium and quotes the real numbers as evidence. No API keys, no third-party service, nothing invented.

perf-a11y-scan.py

Performance · security · a11y

Lab Core Web Vitals (LCP, CLS, FCP, TTFB), an INP lab proxy, page weight, HTTPS/TLS/HSTS/CSP, redirect chain — plus real axe-core WCAG violations and a keyboard focus pass.

mobile-audit.py

Mobile-friendliness

Fit-to-screen, oversized elements, tap targets ≥ 44px and text ≥ 12px, measured per width (320 / 390 / 414) with screenshots.

design-scan.py

Design tokens + templates

The real palette, fonts, and button-style count with desktop/mobile renders; with --pages, interior-template screenshots and measured readability.

seo-sweep.py

Site-wide SEO + schema

A static per-page sweep of the secondary pages with local structured-data validation — pure Python, no browser needed.

tracking-scan.py

Measurement-readiness

Is GA4 / conversion tracking / a real consent layer actually set up? Do trackers fire before consent?

interaction-scan.py

Interaction + links

Does the CTA reach a real destination and does the form validate, across key pages — plus the HTTP status of every link, all without ever submitting.

Graded against 2026. Every “Pass” and every gap is measured against a living benchmarks-2026.md: Core Web Vitals ≤ 2.5s / 200ms / 0.1, tap targets ≥ 44px, WCAG 2.2 AA, unique 50–60-char titles, HTTPS + HSTS, schema, AEO-readiness. The Action Report renders a “2026 standard vs your site” scorecard.
04

Who it’s for

Anyone who owns a site’s outcomes and wants an honest, prioritised read — without a five-figure consulting engagement.

Founder / owner
“Is my site actually working, and what should I fix first?”
→A graded scorecard and a do-this-first roadmap you can hand to a freelancer or do yourself over a weekend.
Marketer / growth
“I need a conversion & SEO teardown I can act on this week.”
→CRO, messaging, and SEO findings with quoted evidence and worked-example fixes, sorted by Impact × Effort.
Agency / freelancer
“I want to open every engagement with a credible audit.”
→A share-ready, print-clean report your client understands — measured where it matters, branded as your process.
Developer / designer
“Give me real Core Web Vitals, WCAG, and mobile numbers.”
→Tool-backed measurements from a real headless render, quoted as evidence — no API keys, runs on Win/mac/Linux.
05

Three ways in

The skills are plain Markdown with YAML frontmatter — version-controllable and editable outside any tool. Pick the path that fits you; all three are free.

Path 1 · zero setup

Copy & paste

Open a skill’s SKILL.md, paste it into a new Claude chat, and send your URL. Works on any plan. Start with Site Snapshot.

Path 2 · recommended

Install as a Skill

In Claude, turn on Settings → Capabilities → Skills, upload each .skill package, and just ask “Audit my website.”

Path 3 · for developers

Use in Claude Code

Drop the unpacked folders into your skills directory and get the measured Extended layer — real renders, timings, and scans.

Extended layer — one-time setup, any OS
# 1 · the environment that can run the tools
npm install -g @anthropic-ai/claude-code

# 2 · the render engine (bundles its own Chromium)
pip install playwright
playwright install chromium

# then, in Claude Code:
"Run the full website audit of https://mysite.com"
Claude launches the measured scans in the background, interviews you while they render, and fills the numbers into the report. No headless browser? Every measured check falls back to not measured — it never guesses.
06

Runs where you already are

Native Claude Agent Skills. Run them on a setup that can fetch web pages — or paste the page text and they work from that. For best results, use a top-tier model.

claude.ai (web) Claude Desktop Claude Code Claude API / Agent SDK Windows macOS Linux no API keys
07

Questions

Is it really free?
Yes — MIT licensed, both layers. The only difference between Basic and Extended is a one-time local tool setup that unlocks measured evidence. You never lose anything by staying on Basic; Extended just fills in hard numbers. No API keys, no third-party service.
Does it make up numbers?
No. What the bundled Python tools can measure locally is measured (Core Web Vitals, accessibility, security, mobile fit, design tokens, links, schema). What only your own analytics can show — conversion rates, traffic, rankings — is owner-reported in the interview or left out, never invented. In Basic, measured checks fall back to “not measured.”
What do I actually get out of it?
Per skill, two files: a .md and a styled, self-contained .html that prints cleanly to PDF. The Action Report bundles everything into one .zip with a single consolidated report as the entry point — reports on top, screenshots in assets/, raw measured JSON in data/.
Do I need to run all eight skills?
No. Start with Site Snapshot (skills 2–6 build on it), then run one, some, or all of the middle five. The Action Report synthesises whatever you ran into one prioritised plan.
Which AI does it need?
These are Agent Skills for Claude. Run them in the Claude web/desktop app (Skills are on paid plans), in Claude Code, or via the API/Agent SDK. Because the suite is URL-first, use a setup that can browse — or paste the page text/screenshots. For best results, use Claude Opus or Sonnet.
Can I white-label it for clients?
Yes. MIT means you can use, adapt, and share it commercially — keep the copyright notice; a visible credit is appreciated, not required. The report’s HTML design system is editable, so you can brand the output as your own process.
An open invitation

Point it at your site. See what a consultant would say.

Free, open source, and honest by design. Fork it, adapt the rubric, add a dimension — or just run it on your homepage tonight.

Get the suite on GitHub →