346 Tokens

Cloudflare recently shipped Markdown for Agents. Any site on their network can serve markdown to agents and HTML to humans from the same URL. Content negotiation at the edge. One toggle in the dashboard. You could build this yourself, manage your own content negotiation layer, serve your own markdown endpoints. This is just the easy mode version.

I turned it on.

What Agents See#

This is my homepage. Dark mode, light mode, testimonial cards, hover states, the works.

The version humans see. Navigation chrome, theme toggle, testimonial cards, footer.

Then I asked for the same page as markdown.

curl -s -D - -o /dev/null \
  -H "Accept: text/markdown, text/html" \
  -H "Accept-Encoding: identity" \
  https://tonyseets.com/

HTTP/2 200
content-type: text/markdown; charset=utf-8
vary: accept
x-markdown-tokens: 346
content-signal: ai-train=yes, search=yes, ai-input=yes

346 tokens. My entire homepage is 346 tokens to an agent.

Here’s what they get:

[Skip to main content](#main-content)

# Relentlessly resourceful.

"These yaks aren't going to shave themselves!"

Status

Not available for projects

Agentic Engineer
Webflow Specialist | Advisor
Explorimenting Web | AI

## What People Say
...
## Schedule a Call
Pick a time that works for you.

That’s it. No hero animation. No grid system. No carefully chosen typefaces. A heading, a tagline, some labels, and a testimonial section.

My name isn’t in there. “Tony Seets” doesn’t appear anywhere in the markdown output of my own homepage.

An agent could infer ownership from tonyseets.com. But the page content should carry that context on its own.

Not Just Two Readers#

I framed this as agents versus humans in my last post. That’s too clean.

The markdown layer gets read by LLMs, search crawlers, accessibility tools, and browser reader mode. Each one strips the page down in a different way. I’ve been designing mostly for the reader that renders pixels.

The Gap#

The homepage markdown is structurally thin because the page relies on visual design to carry meaning.

The punchline is simple: my name doesn’t appear in the markdown output of my own homepage. An agent reading it wouldn’t know whose site this is.

Compare that with my about page. Same site, same toggle.

# About

Not a traditional dev or a pixel pusher, but a builder.
I enjoy working a problem until it gives. My focus is
on the mechanics of turning "what if" into working reality.

For the better part of a decade I've done that in the
Webflow and no-code space...

That page has substance in the source itself. A bio, a point of view, identity. The homepage markdown still reads like a layout template because that’s what it is.

I also found a policy mismatch while testing. robots.txt was blocking several AI crawlers while the content-signal header said ai-train=yes, search=yes, ai-input=yes. That is now aligned in this repo so both signals tell the same story.

Flipping the toggle was easy. Making the content worth serving is the actual work.

The Actual Work#

The toggle gave agents access. Now I need to give them something worth accessing.

This feels closer to the primitives. Less GUI theater, more getting the bones right. Cut fluff. Make sure the core content survives without CSS, layout tricks, or animation.

The crawl policy was the obvious starting point, and it is now fixed. robots.txt and content-signal now tell the same story. Next is thinking about structured orientation for machine readers. llms.txt is one approach, content negotiation is another, structured data is another. Sitemaps list URLs. Orientation layers tell readers what matters and where to start.

I don’t think this makes the web less exciting. It probably makes it wilder for a while. More experimentation, more creativity, more security mistakes, more fast corrections. Wild West energy.

I’m most curious about what happens when every site has a markdown layer. Which pages hold up, which pages collapse, and how fast teams adapt.

The real work isn’t the toggle. It’s content worth serving.