llms.txt generator

What this llms.txt generator can prove

An llms.txt file is an emerging Markdown convention for giving an agent a concise route into a site. Version 2 of the proposal was published in August 2026. It requires only one H1 project or site name. A blockquote summary, details, and H2 file lists are optional, though a file with no useful links is little more than a label. Each file-list row has a linked name and may add a short note explaining what the target contains.

The format check on this page parses the generated Markdown independently. It can prove heading order, link-row grammar, safe public URL syntax, LF line endings and the exact UTF-8 byte count. It cannot prove that a linked page exists because the tool makes no network request. It also cannot prove that an AI product finds, reads, trusts or cites the file. Chrome Lighthouse currently calls the convention emerging and treats a missing file as not applicable, not as a failed requirement.

Why a sitemap does not become a set of invented summaries

A normal sitemap contains locations. It usually does not contain the page title, purpose, audience or a reliable description. When you paste or upload sitemap XML, this generator imports the exact <loc> values and proposes an editable label derived from each URL path. The source badge stays visible. Notes are left empty for you to write from the actual page rather than being filled with language that merely sounds plausible.

Sitemap indexes are handled with the same boundary. Their locations point to child sitemaps, not content pages, so those rows are labeled as child sitemaps and remain unfetched. Paste or upload a child file explicitly if you want its page locations. Namespaced XML, entity-encoded ampersands and CDATA are read locally; malformed XML gets a concrete structural error instead of a partial, apparently successful list.

No website crawler is hiding behind the URL field

The leading generators in this search category ask for a domain and crawl it on a server. That creates two problems. It obscures which source produced each sentence, and arbitrary server fetching creates an SSRF boundary around private addresses, DNS changes and redirects. A browser-only fetch is not a clean substitute because most sites do not grant another origin permission to read their pages.

This route therefore accepts material you can inspect: structured fields, a local sitemap, pasted sitemap XML or one URL per line. Nothing is fetched. If you need clean Markdown versions for linked HTML pages, the HTML to Markdown converter handles supplied page source without pretending it crawled the live site.

Publish the concise guide, not a disguised site archive

Version 2 expects agents to read or search the concise file, choose a relevant link, and fetch detail only when needed. That is different from the informal llms-full.txt pattern, which concatenates page content into a large artifact. Full-content bundles have separate freshness, copyright and context-size risks and are not generated here. A sitemap dump is not much better: search engines already have sitemaps, while this file is supposed to curate the pages that matter.

Prefer a few sections with names a reader can understand. Link to clean Markdown or text versions when your site publishes them; version 2 recommends those targets but does not make them a validity requirement. The checker reports an HTML-looking target as advice, not an error, for exactly that reason.

Install and maintain the downloaded file

Download llms.txt and publish it at the path it describes, commonly the site root. Version 2 also permits a file below a path such as /docs/llms.txt, where it covers that subtree; the most specific file applies. The proposal additionally recommends standard rel="describedby" links or HTTP Link headers to identify the covering file, and rel="alternate" type="text/markdown" for clean page alternatives.

Regenerate when the curated structure changes, then review the source badges and notes before replacing the published file. The output is deterministic: the preview, clipboard and download use the same LF-only text with one final newline. No account, crawler, model call or upload is involved.