Most AI integrations with websites rely on the visible surface: the DOM. AI scrapes. It clicks. It simulates users. And we accept fragility as inevitable.

In this talk, I present a new method for building AI Skills (or MCP servers) that bypasses brittle DOM automation entirely. During the development phase, instead of scripting what the user sees, we use Playwright to observe what the front-end calls. By programmatically browsing the application and capturing its backend API traffic, we can automatically generate an API document — effectively reverse-engineering the site's operational contract.

Once that API document exists, the integration changes entirely. We write the AI Skill directly against the backend API. No scraping. No selectors. No fragile page structures. Just structured requests and responses — the way the application was meant to operate. In short, this method reframes websites not as pages to automate, but as APIs waiting to be documented.

If AI agents are going to operate the Internet, they shouldn’t behave like distracted humans clicking buttons. They should behave like first-class clients of the system.

This talk demonstrates the architecture, tooling, and workflow behind this approach — and argues that the future of AI integrations lies beneath the interface, not on top of it.

You will learn:

• How to extend AI beyond the built-in tools and libraries
• Capitalize on all the websites so that you get stuff done (not just answer questions)
• How to create skills for AI