Skip to content

kraayenjon/claude-pagekit / skills/llms-txt

llms-txt

Verified

Write and validate an llms.txt file — the /llms.txt convention that gives AI assistants a clean, curated map of a site instead of leaving them to guess from HTML. Covers the required format, what belongs in it and what does not, the optional llms-full.txt companion, and where to put both. Use when the user says "llms.txt", "llms-full.txt", "AI crawler file", "help AI understand my site", "make my site readable to ChatGPT", or is preparing a site to be cited by AI assistants.

llms.txt

A single Markdown file at /llms.txt that tells an AI assistant what a site is and where the good pages are. It exists because HTML is noisy — navigation, cookie banners, footers, scripts — and a model reading a page has to guess what matters. llms.txt removes the guessing.

It is a convention, not a standard. No crawler is required to read it. Write one anyway: it costs ten minutes, it is read by a growing number of assistants and tools, and the act of writing it forces you to state plainly what the site is for.

The format

Strict Markdown, in this order. The structure is the contract — a parser looks for exactly this shape.

# Site or business name

> One sentence saying what this is. Concrete. No marketing adjectives.

Optional free prose. A short paragraph or two of context that does not fit a link —
what the business does, who it serves, where it operates, what makes it different.
Keep it factual.

## Section name

- [Page title](https://example.com/page): what a reader gets from this page
- [Another page](https://example.com/other): one clause, not a sentence

## Optional

- [Lower-priority page](https://example.com/archive): things a model can skip if short on space

Rules that actually matter:

  1. One # H1, the name. First line of the file.
  2. A > blockquote directly after it, one sentence. This is the line most likely to be quoted back verbatim, so write it like a caption, not a slogan.
  3. ## sections group the links. Name them for what they are — Services, Case studies, Docs, About — not Resources or More.
  4. Every list item is a link followed by : and a description. The description is the point. [Pricing](/pricing): three plans, from €19/month, no setup fee is useful. [Pricing](/pricing): our pricing page is noise.
  5. Absolute URLs. A model may read the file with no knowledge of the origin.
  6. A section literally named ## Optional is the one piece of magic: anything under it is explicitly marked as skippable when context is tight. Use it. Do not use it for anything you actually want read.

What goes in

Only pages a person would be glad to land on. A good llms.txt for a small site is ten links. For a large one it is thirty, not three hundred.

Include: the home page, what you sell or do, pricing, about, contact, the three or four best pieces of writing, documentation entry points, anything with real specifics in it.

Leave out: tag and category archives, pagination, login and account pages, checkout, thank-you pages, anything noindex, anything thin, and anything duplicated across URLs.

The one-page case

A single landing page still gets an llms.txt, and it is mostly prose rather than links. The blockquote and the paragraph under it carry everything:

# Panadería Aurora

> A wood-fired sourdough bakery in Malasaña, Madrid, open Tuesday to Sunday from 8am.

Aurora bakes four breads daily — country sourdough, seeded rye, olive fougasse and a
Saturday-only brioche — in a wood oven installed in 1974. Everything is sold from the
counter at Calle del Pez 14; there is no delivery and no online ordering. Wholesale
enquiries from restaurants are taken by phone.

## Pages

- [Home](https://panaderiaaurora.es/): breads, hours, address, and how to order wholesale

## Optional

- [Instagram](https://instagram.com/panaderiaaurora): daily bake announcements

That is a complete, correct file. Do not pad it.

llms-full.txt

The optional companion at /llms-full.txt holds the actual content — the full text of the important pages, concatenated as clean Markdown, so an assistant can read the substance without fetching anything.

Write one when the site is documentation, a knowledge base, or anything where the value is in the text. Skip it for a one-page landing site: llms.txt already contains everything.

Keep it plain Markdown: # Page title, then the page's prose. No navigation, no HTML, no scripts, no repeated header and footer. Strip everything a reader would skip.

Where the files go

Both live at the domain root, served as text/plain or text/markdown:

FrameworkLocation
Next.js (App Router)public/llms.txt
Astro, Vite, SvelteKitpublic/llms.txt
Hugo, Jekyll, Eleventystatic/llms.txt
Plain static sitenext to index.html

Verify after deploying: https://yourdomain.com/llms.txt must return the raw file, not an HTML 404 page.

Then point robots.txt at it, so a crawler that does not know the convention still finds it:

User-agent: *
Allow: /

User-agent: GPTBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Google-Extended
Allow: /

Sitemap: https://yourdomain.com/sitemap.xml

llms.txt is not a robots.txt directive, so it has no line of its own there. Linking it from the home page <head> is harmless and occasionally helps:

<link rel="alternate" type="text/markdown" href="/llms.txt" title="llms.txt">

Validating one

Read it back and check:

  • Exactly one # heading, and it is the first line
  • A > blockquote immediately after it, one sentence, factual
  • Every ## section has at least one item
  • Every list item is - [text](absolute-url): description
  • No relative URLs
  • No link to a page that 404s, redirects, or is noindex
  • Every description says something specific — no "our services page"
  • Under about 100 links; if longer, most of them belong under ## Optional or nowhere
  • Nothing in it is untrue. A model will repeat this verbatim

The last one is the important one. llms.txt is the most quotable file on a site — everything in it is written in a model's preferred format, pre-chewed and unambiguous. An invented claim here does not get buried the way it would in a paragraph of HTML. It gets repeated.

Writing the blockquote

It is one sentence and it does most of the work. Three failure modes:

BadWhyBetter
"Innovative solutions for modern businesses."Says nothing. Fits any company."Payroll software for Spanish companies with 5–50 employees."
"The best bakery in Madrid."Unverifiable claim, reads as marketing, gets discounted."A wood-fired sourdough bakery in Malasaña, Madrid."
"We are a full-service creative agency offering branding, web design, motion graphics, packaging and strategy for clients across sectors."Too long, no shape, nothing to lift."A three-person branding studio in Lisbon working mostly with restaurants and hotels."

Name the thing, the who, and the where. Numbers and place names survive summarisation; adjectives do not.

slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall