A community convention for a markdown file at /llms.txt that gives a large language model a curated, link-rich map of a site in a form it can actually consume — proposed by Jeremy Howard in September 2024. It is not an IETF or W3C standard and has no registry behind it; it spread because the problem it names is real and nothing else was addressing it.
llms.txt
llms.txt is a markdown file served at the root of a site that tells a large language model what the site contains and where the good parts are. It exists because the alternative — pointing a model at HTML built for humans, wrapped in navigation, cookie banners and ad slots — wastes most of a context window on markup before it reaches a sentence worth reading.
- Markdown, not a new syntax - An H1 with the site name, a blockquote summary, then link lists under H2 headings. Anything that can write a README can write one.
- Curated rather than exhaustive - It is a map, not a sitemap. The point is to name the documents that matter and skip the ones that do not.
llms-full.txtas the companion - Wherellms.txtlinks out, the full variant inlines the content, trading size for a single fetch.- No registry, no conformance - There is no body that blesses one, no validator that rules on one, and no mechanism that makes a site serve one. Adoption is the only thing holding it up.
I track it because it is the clearest example of the pattern I keep writing about: an unclaimed slot
between systems gets filled by whoever writes the document and convinces people to use it. llms.txt
was not standardized into existence — it was proposed, and it spread. Whether it survives contact
with the AGENTS.md family, MCP and
.well-known discovery is genuinely open, and the honest position is that it is
a convention with momentum rather than a standard with authority.
The question I get asked most about it is answered separately in the FAQ: do I need an llms.txt?
