On my own

Built for agents

Ongoing · part of this site

Problem: More and more of the web is read by machines acting for a person: research assistants, coding agents, and the tools that summarize or cite a page before anyone clicks it. An agent lands on a page, tries to pull meaning out of a wall of markup, and often comes away with a fuzzy version of it. I didn’t want an agent’s read of me to be a scrape of my HTML and a guess. If that reading is going to happen anyway, the site should offer a clean version of itself on purpose, instead of forcing a machine to reverse-engineer one.
Constraints: One source of truth, no second copy to keep in step. Whatever an agent reads has to come from the same pages a person reads, or it drifts the moment I edit. HTML stays the default, because browsers are still the main audience. And the machine version had to cost nothing to maintain, or I wouldn’t keep it honest.
Process: I built it as a layer over the normal site. The tempting shortcut was to hand-write a Markdown copy of each page, but two copies drift the moment you edit one, so I ruled it out. Instead, after each build a script takes every rendered page, lifts the main content, and generates a Markdown twin, so the agent version is derived from the real page and can’t fall behind it. Middleware reads the request: ask for HTML and you get HTML, prefer Markdown and you get the twin at the same URL. A header on the home page points agents at a curated llms.txt and my profile, and structured data spells out who and what the site is. Studio and Notebook both read from one data file, so a page, its index entry, its sitemap line, and its Markdown twin all trace back to the same source.
Outcome: The site now answers both audiences from a single build. Send a normal request and you get the page. Ask for Markdown and you get a clean, structured version of the same content, no scraping required. It stays correct on its own because it is generated, not written by hand. It is a small piece of infrastructure, and it is the kind I like: it does its job without asking for attention.
Lessons: Building with AI changed how I work. Building something for AI to read changed what I make. Once you treat an agent as a real reader, the questions get concrete: what it should see first, what it can skip, and how to keep the machine version honest without maintaining it by hand. It is the same product work I already do for people.

Verify

This one you can check rather than take on faith. Read the curated guide for agents, or ask any page on this site for its Markdown twin.

llms.txt ↗this page as Markdown ↗

curl -H "Accept: text/markdown" https://bennieiv.xyz/studio/built-for-agents

This website →

← Back to Studio