llms.txt

The llms-txt generator writes an llms.txt file: a Markdown index of your docs that links to each page with a one-line description, so language models and AI tools can find their way around. Add it when you publish your docs on the web.

This page also covers llms-full.txt (llms-txt-full), which puts the Markdown of every page in one file.

Target: llms-txt · Package: @doc-kit/generator-react · Reads: API entries · Writes: llms.txt

The links in llms.txt are absolute URLs, so set baseURL to where your docs are published. There is no flag for it, so set it in your configuration file:

export default {
  global: {
    baseURL: 'https://example.com',
  },
};

npm install --save-dev @doc-kit/cli @doc-kit/generator-react
npx doc-kit generate -t llms-txt -i "docs/**/*.md" -o out

For a project named my-lib with docs/index.md and docs/guide/usage.md, this writes:

# my-lib Documentation

> API documentation for my-lib.

Below are the sections of the API documentation. Look out especially towards the links that point towards guidance/introduction to the structure of this documentation.

## API Documentations
- [Usage](https://example.com/guide/usage.md): How to call my-lib from your code.
- [My library](https://example.com/index.md): A small library for working with strings.

Options go in the llms-txt section of the configuration file:

import { join } from 'node:path';

export default {
  'llms-txt': {
    templatePath: join(import.meta.dirname, 'llms-template.txt'),
    pageURL: '{baseURL}/docs{path}.md',
  },
};

Like every generator section, it can also override a global option, such as baseURL or project, for this generator alone.

llms.txt has one line for each top-level (#) heading in your sources, as - [Heading](URL): Description:

• Heading is the heading's text. A page without a # heading isn't listed, and a page with several gets a line for each.
• URL is pageURL filled in for the page. {path} is the source file's path relative to the base of the input glob: with -i "docs/**/*.md", docs/guide/usage.md becomes /guide/usage.
• Description is the page's llm_description when it has one. Otherwise it's the first paragraph between the heading and the next heading, on one line: inline code, bold and italics keep their Markdown, and links become plain text. Without such a paragraph, it's empty.

If your opening paragraph doesn't make a good summary on its own, write one for models with the llm_description directive, right after the heading it describes:

# Usage

<!-- llm_description=How to call my-lib from your code. -->

Section pages aren't listed: llms.txt links to the full pages.

llms.txt is the template's text followed directly by the list of links, so end the template with a newline. This is the default:

# {project} Documentation

> API documentation for {project}.

Below are the sections of the API documentation. Look out especially towards the links that point towards guidance/introduction to the structure of this documentation.

## API Documentations

The format at llmstxt.org expects a # title with your project's name, then a > summary, then ## sections of links. A template of your own can describe your project better and add links of its own above the generated list.

Placeholders take the value of that option from the generator's configuration, which includes every global option:

A placeholder without a value is left as written. To use a fallback instead, write {name|fallback}, where the fallback is a single word (letters, digits and underscores). The same placeholders work in pageURL, along with {path}.

The default pageURL links to a .md file at each page's path (/guide/usage.md next to /guide/usage.html), as llmstxt.org proposes, since Markdown is easier for language models to read than HTML. doc-kit doesn't write those files, so copy your sources into the output. With the website, its pathsToCopy option does it:

export default {
  target: ['html', 'llms-txt'],
  global: {
    input: ['docs/**/*.md'],
    output: 'out',
    baseURL: 'https://example.com',
  },
  html: {
    // Copies the contents of docs/ into out/
    pathsToCopy: ['public', { docs: '.' }],
  },
};

Setting pathsToCopy replaces its default list (assets, public and static), so keep the ones you use, as public is kept here. doc-kit's own site builds its llms.txt this way. If you don't serve Markdown, link to the HTML pages instead with pageURL: '{baseURL}{path}.html'.

The llms-txt-full generator writes the Markdown of every page into one llms-full.txt file, each page after its URL, so a language model can read all of your docs in a single request. Pair it with llms.txt, which only links to each page.

Target: llms-txt-full · Package: @doc-kit/generator-react · Reads: API entries · Writes: llms-full.txt

Set global.baseURL as for llms.txt, since each page's URL starts with it, then run:

npx doc-kit generate -t llms-txt -t llms-txt-full -i "docs/**/*.md" -o out

Options go in the llms-txt-full section of the configuration file:

The default URLs point at each page's Markdown file, which none of the built-in generators write: copy your sources into the output, as in Serving pages as Markdown, or set pageURL to '{baseURL}{path}.html' to point at the website's pages.

Every page read from your sources, one after another. Each starts with a url block between --- lines, built from pageURL, followed by the page's Markdown. For this page and a path.md next to it:

# File system

<!-- YAML
added: v1.0.0
-->

The `fs` module reads and writes files. Paths can be built with
[`path.join()`][join].

### `fs.readFile(path)`

- `path` {string|URL} The file to read.
- Returns: {Promise<Buffer>}

[join]: ./path.md#pathjoinparts

llms-full.txt holds:

---
url: https://docs.example.com/fs.md
---
# File system

The `fs` module reads and writes files. Paths can be built with
[`path.join()`](./path.html#pathjoinparts).

### `fs.readFile(path)`

* `path` {string | URL} The file to read.
* Returns: {Promise<Buffer>}

---
url: https://docs.example.com/path.md
---
# Path

### `path.join(...parts)`

Joins path segments.

The Markdown is rebuilt from the API entries, not copied from your files, so it differs from the sources in a few ways:

• Links to .md files point to the .html pages.
• Reference-style links become inline links, and their definitions are dropped.
• Formatting is normalized: bullet lists and emphasis use *, strong text uses **, and a page without a heading starts with an empty #.
• <!-- YAML --> comments are left out.
• Type annotations are rewritten with spaces around | ({string|URL} becomes {string | URL}). MDX pages keep their JSX.