For your LLM

Half the readership of this site is a language model writing YAML on someone’s behalf. So every page is published twice, from one source.

URL	What it is
/docs/…	The rendered page — navigation, search, table of contents
/docs/….md	The same content as plain Markdown, with front matter and no chrome
/llms.txt	An index: one line per page, pointing at its .md twin
/llms-full.txt	Every page concatenated, for a model with room for the lot

Add .md to any docs URL in this site and you get the Markdown. Every page also carries a Copy for chat button that puts exactly that text on your clipboard.

The front matter is the useful part

---
title: "How checks run"
url: https://example.com/docs/validation/how-checks-run
schema: "0.1"
source: packages/validate/src/index.ts
---

schema is what makes a stale copy detectable rather than mysterious: if the material your model is working from names an older schema than the page it is reading, that is a fact either of you can act on.

source names the file in the codebase a page describes — packages/spec/src/anchors.ts for the anchors page, packages/catalogue/data/fixtures/bed-bolt.yaml for the bed bolts. It is a pointer to the authoritative copy of the numbers on the page, and to the file whose change would make the page stale.

generated: true is the stronger claim: that page is written by a script from its source and is never edited by hand, so fixing one means fixing the data. Only the error-code reference carries it today.

If you are an agent reading this

The static half of this site is prerendered HTML: it needs no JavaScript, and a plain HTTP GET returns the whole page. /app/ is the opposite — a client-rendered compiler whose HTML is an empty element — so there is nothing there for you, and robots.txt says so.

The error-code reference is generated from the registry the validator itself reads, so a code that exists and a code that is documented cannot diverge. The rest of the reference — schema and catalogue — is written by hand, because most of what makes those pages useful is not in the data. The authoritative copy of any number on them is the one the validator reads, and a finding will always carry it.

If you want the numbers without the prose, take catalogue.json — the same machine-readable file the authoring pack ships.