Web Anatomy

Skills for your AI agent that ship landing pages that work. Built from the plays the best SaaS companies actually shipped, grounded in a scored benchmark library so every recommendation cites a real page instead of generic advice.

$ npx skills add GabrielAmz/web-anatomy

Installs the Web Anatomy skill pack for Claude Code, OpenAI Codex, Cursor, Windsurf, and any agent that supports the Agent Skills spec.

What Web Anatomy is

Most AI agents rework a page from generic advice: add social proof, make the CTA stronger, tighten the headline. Web Anatomy gives the agent evidence instead. The skills read a page as a structure of known sections (hero, value proposition, trust, testimonial, pricing, FAQ, CTA), then ground every move in a benchmark library of real pages scored section by section.

The benchmark behind the skills (numbers to confirm before release):

• 3,500+ scored sections
• 500+ landing pages
• 290+ companies

The skills run standalone with built-in patterns. Connecting the hosted MCP (highly recommended) pulls the live benchmark data and real reference screenshots into every report. See Live benchmark data.

The four altitudes

Web Anatomy works at four zoom levels. The altitude matches the question being asked.

After the copy is written, build-page assembles the structure + copy into a shareable wireframe with a coherent, single-exemplar design system. It is the terminal step on both branches (improve and create), so it sits above the altitudes rather than inside one.

Run webanatomy-setup once first. It captures product, ICP, industry, competitors, and proof assets into .agents/webanatomy-context.md, so every skill gives sharper, on-brand recommendations.

Two flows

• Improve an existing page: audit-page (diagnose) -> write-page (rewrite the weak sections) -> build-page (optional: assemble the improved page as a shareable wireframe, in your page's own look).
• Create a new page: find-examples (pick one exemplar homepage) -> write-page build mode (a light outline adopts the exemplar's section order, then writes the copy) -> build-page (assemble the wireframe in the exemplar's look).

The skills

Visual Reports

Web Anatomy does not rely on chat to render screenshots. The output skills write local visual artifacts:

.webanatomy/
├── find-examples/
├── audit-page/
├── research-best-practices/
├── moodboard/
│   └── page-date/
│       ├── report.md
│       ├── report.html
│       └── references/
├── write-page/
│   └── topic-date/
│       ├── report.md
│       ├── report.html
│       └── references/
│           ├── current.png
│           └── company-section.png
└── build-page/
    └── page-date/
        ├── wireframe.html
        ├── design-system.md
        └── references/

The chat response should stay short: top findings, report path, and any screenshot or MCP limitations. report.html is the product experience.

Live benchmark data (optional, highly recommended)

The skills work on their own, so install them and start immediately. Connecting the hosted Web Anatomy MCP is never mandatory, but highly recommended: it grounds every recommendation in live benchmark data and real reference examples instead of static guidance, which is where the skills are at their best. It needs a beta token and one config command per IDE:

Sign in at https://www.webanatomy.ai/dashboard/mcp to get your token + per-IDE config. You must be signed in, because the page generates your personal token there.

The skills use MCP tools such as search_pages, get_page, search_sections, and get_section when connected, and degrade gracefully to static guidance when not. Scores, thresholds, raw marker coordinates, and benchmark field names stay internal; user-facing reports translate them into plain-English patterns and gap labels.

Want to see the benchmark first? Browse the public scored sections gallery at https://www.webanatomy.ai/best-landing-pages/sections — real landing-page sections, scored, with what makes each one work. It is the same library the MCP serves to your agent, so you can explore it before connecting anything.

Why "Anatomy"

Most AI-agent skill packs frame pages as a job (CRO, copywriting, SEO). Web Anatomy frames them as a structure: hero, proof, problem, solution, pricing, FAQ, CTA. Every section has known patterns, known anti-patterns, and known reference companies.

Every Web Anatomy skill turns a page into parts: hero, proof, problem, solution, pricing, FAQ, CTA. Each part has known patterns, anti-patterns, and benchmark examples. The framework and the evidence travel together.

Compatibility

• ✅ Claude Code
• ✅ OpenAI Codex
• ✅ Cursor
• ✅ Windsurf
• ✅ Any agent that reads .agents/skills/<name>/SKILL.md per the Agent Skills spec

Install options

Option 1 — CLI (any agent):

# Installs the whole pack to .agents/skills/
npx skills add GabrielAmz/web-anatomy

# Update later (or just re-run the command above, it's idempotent)
npx skills update GabrielAmz/web-anatomy

Option 2 — Claude Code plugin (version-aware updates):

/plugin marketplace add GabrielAmz/web-anatomy
/plugin install web-anatomy

Claude Code knows the installed version and when a newer one ships. Update with /plugin marketplace update web-anatomy then /plugin update web-anatomy, or turn on auto-update for the marketplace to make it hands-off. Skills install namespaced: /web-anatomy:write-page, /web-anatomy:build-page, and so on.

Option 3 — clone and copy manually:

git clone https://github.com/GabrielAmz/web-anatomy
cp -r web-anatomy/skills/* .agents/skills/

After install, start with webanatomy-setup. The benchmark-backed workflows degrade gracefully when MCP isn't connected, falling back to static/reference guidance, but they are at their best with live benchmark data, so connecting MCP is highly recommended.

Repo layout

skills/ holds the skill pack, the source of truth installed by npx skills add.

Contributing

PRs welcome. See CONTRIBUTING.md for the new-skill recipe: kebab-case, YAML front matter, clear workflow names, and body under 5K tokens.

License

MIT. Use it, fork it, sell it, ship it.

Built by Gabriel Amzallag. The Web Anatomy pack is a public extraction of the patterns powering a private landing-page benchmark and audit tool. The skills give you the framework; the analyzer gives you the data.