How this site is built
We build AI systems for clients, so here is how we built our own site, and what we checked before calling it done.
- Built with
- Astro, static HTML
- Served by
- Cloudflare Workers
- Type
- Schibsted Grotesk, IBM Plex Mono
- Drawings
- SVG, generated from data
- Every page as Markdown
- Add index.md
- For agents
- llms.txt, rates.json
Words in markdown, facts in one catalog
Every page starts as a markdown file that people can read and edit. Every price, term, and booking link comes from one catalog file. If a page and the catalog disagree, the build uses the catalog and prints a warning, so a price can’t drift between the homepage and a program page.
Static pages, almost no JavaScript
The site is built with Astro into plain HTML and served by a Cloudflare Worker. The program picker on the homepage works with CSS alone. Fonts are self-hosted, with fallbacks measured to match them so the page doesn’t jump when they load.
The drawings
The isometric figures are generated as SVG by a short script from the same data the pages use: the seven plates in Fig. 0 are the seven domains the AI Audit scores. There is no stock illustration and no AI-generated imagery on the site.
For AI agents
- Every page is also available as Markdown. Add
index.mdto its address, or request it withAccept: text/markdown. - /llms.txt lists the pages, and /llms-full.txt has the whole site in one file.
- /rates.json has every program’s price and terms as data.
- Each page carries JSON-LD that matches what’s visible on it: the organization, the people, the programs and their prices, and the Summit.
- The homepage program finder is a plain HTML form with WebMCP attributes, so a browser agent can use it as a tool.
Built with AI agents, checked by people
The site was written with Claude Code from a written design system and the services copy. Dan reviews every page, and every build is checked for leftover author notes, unconfirmed claims, and Lighthouse scores.
Updated