Publish llms.txt. It is the only file the llmstxt.org proposal defines. llms-full.txt is not in the specification text, so it is an ecosystem convention that you can add if you have a reason to [5][6].
This article walks through what the proposal says, what a valid file looks like, where a "full" file fits, and what the proposal does not claim.
What the proposal defines
The proposal is titled "The /llms.txt file, v2." It was published on September 3, 2024, and modified on August 10, 2026 [1]. It describes a small Markdown file that summarizes a site and points to more detailed pages.
The point that matters most for this comparison is that the specification text does not mention llms-full.txt [5]. An independent explainer reaches the same conclusion: the spec defines llms.txt and linked Markdown versions of pages, not a full file [6].
The llms.txt format
The format is Markdown with a fixed order of parts [1]:
| Part | Required? | Purpose |
|---|---|---|
| H1 title | Yes, the only required part | Names the site or project |
| Blockquote | Optional | Short summary of the site |
| Detail sections | Optional | Plain paragraphs or lists with context |
| H2 sections of links | Optional | Lists of Markdown links to deeper pages |
The first line is an H1 heading, written as a # followed by your site name. A blockquote summary comes next, then any detail text, then H2-delimited lists of links. Here is the part of a file that follows the title:
> Acme makes invoicing software for freelancers. This file lists our docs and policies.
Pricing is per seat. All docs are available as Markdown.
## Docs
- [Quick start](https://acme.example/docs/quick-start.md): Create your first invoice
- [API reference](https://acme.example/docs/api.md): Endpoints and auth
## Policies
- [Terms](https://acme.example/terms.md): Terms of service
Each link should point to LLM-friendly content, such as a Markdown version of a page [1]. The file itself is meant for agents to view or search before they follow links [1].
For a step-by-step build, see how to write an llms.txt file. For background on why the file exists, see what llms.txt is.
Where the file can live
Version 2 allows an /llms.txt file at the site root or at any subpath. It covers the pages under that path [1]. When more than one file matches a path, the most specific one applies [2].
That matters for large sites. A docs section at /docs/ can have its own file without you cramming everything into one root file.
Discovery: link relations
Version 2 recommends two link relations [1]:
rel="alternate"withtype="text/markdown"marks the Markdown version of a page.rel="describedby"marks the llms.txt file that covers a page.
You can deliver either one through an HTML link element in the page head or through an HTTP Link header [1]. The proposal gives a standard way for a tool to find the Markdown and the index, so you don't depend on a crawler guessing URLs.
The Optional section changed in v2
Version 1 gave a section named Optional a special mechanical meaning. Version 2 removed it. Optional sections are still allowed, but only as a convention for secondary links [2]. If an older tutorial says a tool will skip that section to save context, the current text no longer says that.
So where does llms-full.txt fit?
The proposal's design is a compact index with the detail behind links, fetched only when needed [1]. A single file holding all your content is a different design. The name comes from the ecosystem, not the spec [5].
In our reading of how sites use it, the full file is a bundle of the content behind the links in one place, so a tool can fetch everything with one request. That is common practice, not something the proposal defines. We have no vendor documentation in hand saying a given AI engine fetches either file, so treat any benefit as unconfirmed.
| llms.txt | llms-full.txt | |
|---|---|---|
| Defined by llmstxt.org | Yes [1] | No [5][6] |
| Required H1, optional rest | Yes [1] | No format to follow |
| Role | Index: summary plus links | Convention: bundled content |
| Size guidance | Small enough to fit in context [1] | None from the spec |
Size considerations
The proposal says the file should stay small enough to fit in context, with detail living behind linked pages [1]. It gives no byte or token limit in the notes we have, so don't quote one.
Practical consequences:
- Keep llms.txt to a summary and curated links. If it reads like a sitemap dump, trim it.
- A full file has no spec-backed size target. If you publish one, it grows with your content, and it may exceed what a given tool can load. Test it with the tools your audience uses.
- For a large site, prefer several path-scoped llms.txt files over one giant file [1][2].
- A full file is another copy of your content to keep in sync. If you can't automate that, skip it.
What the proposal does not claim
- It does not define
llms-full.txt[5]. - It does not present the file as a ranking signal or a training control. It describes the file as something agents view or search before following links [1]. That is our reading of the text, not a line it states.
- It does not replace existing standards. The proposal says llms.txt coexists with them and complements sitemaps and robots.txt [3]. Crawl access still belongs in robots.txt, and URL discovery still belongs in your XML sitemap.
- It does not promise citations or traffic.
Generation support exists. The proposal's integrations section lists Mintlify, GitBook, Yoast SEO, AIOSEO, and Wix as producing llms.txt automatically [4]. If you use one of them, check what it emits before writing by hand.
What to do next
- Publish a valid
/llms.txt: an H1, a short blockquote, and H2 lists of links to Markdown pages. If you want a draft, the free llms.txt generator builds a spec-shaped one from a live homepage and sitemap. - Serve Markdown versions of the pages you link to, and add the
alternateanddescribedbyrelations. - Add
llms-full.txtonly if you can generate it automatically and have a tool that uses it. Label it in your docs as a convention, not a standard. - Leave crawler rules in robots.txt and your URL list in your sitemap.