Agent-Friendly Web

The llms.txt format used here

What this site's llms.txt and llms-full.txt contain, in what order, and how the build generates them.

Updated

This site publishes /llms.txt following the llmstxt.org proposal, and a companion /llms-full.txt. For the general convention and why it helps agents, read the guide entry on the llms.txt file.

Structure of llms.txt

The file is Markdown, served as text/plain; charset=utf-8, and is kept under 4 KB. Its parts always come in this order:

  1. An h1 with the site name. This is the only part the spec requires.
  2. A blockquote with a one-sentence summary of the site.
  3. One paragraph explaining the URL conventions: append index.html.md for Markdown, and use /api/v1/pages/{id}.json for JSON.
  4. ## Docs: every page in this docs section.
  5. ## Guide: every guide entry.
  6. ## Machine endpoints: the JSON index, the three schemas, the feed and the sitemap.
  7. ## Optional: links an agent can skip when short on context, namely llms-full.txt and the about page.

Every list item has the form - [Title](absolute URL): summary. Page links point at the Markdown copy, not the HTML, so an agent gets clean text in one request. There are no headings deeper than h2.

Example

The start of the generated file looks like this:

# Agent-Friendly Web

> A small, fully static guide to web conventions that make sites easy for LLM agents to read, cite and crawl.

Append `index.html.md` to any page URL for Markdown. ...

## Docs

- [URL and anchor policy](https://agentsboard.duckdns.org/docs/url-policy/index.html.md): The URL scheme, ...

llms-full.txt

/llms-full.txt contains the Markdown copy of every page in site order, separated by a line containing only ---. Each copy starts with # Title, the summary as a blockquote, and Canonical: and Updated: lines, so an agent can split the file and still attribute each part. The spec does not define this file; it is a common addition for agents that prefer one large fetch.

How it is generated

build.py writes both files from the same page list it uses for HTML, so they cannot drift out of date. Titles and summaries come from each page's front-matter, and every URL is built from base_url. validate.py then checks the order of the parts, the item format, that every link is absolute and resolves to a built file, that an ## Optional section exists, and the 4 KB limit.