llms-txt-doctor

llms-txt-doctor

Diagnose and generate llms.txt files. One command, zero config, zero dependencies, no API keys.

# Check a site's llms.txt — spec lint, dead links, sitemap coverage
npx llms-txt-doctor check example.com

# No llms.txt, or a broken one? Generate it
npx llms-txt-doctor generate example.com

Why another llms.txt tool?

Most generators stop at generating. In practice the problems we kept seeing are after that: files that return HTML because of SPA fallback routes, dead links, an llms.txt that covers 5% of the site and never gets updated. So this tool treats checking as a first-class command, not an afterthought:

• ✓ Spec structure: H1 title, > summary, ## sections, link items
• ✓ Serving problems: 404s, HTML bodies, wrong content-type
• ✓ Link liveness (sampled, polite)
• ✓ Coverage vs sitemap — what % of your site does your llms.txt actually list?
• ✓ Bilingual gap — site has Chinese (/zh/) pages but llms.txt lists none
• ✓ llms-full.txt presence

And it's honest about scope: generate builds the index from your sitemap (or homepage links as fallback) using titles + meta descriptions. If you own the site's source, a build-time plugin like vitepress-plugin-llms will produce richer content — use that. This tool is for sites you don't have the source of, quick starts, and audits.

Usage

generate <url>

npx llms-txt-doctor generate example.com
npx llms-txt-doctor generate example.com --include /docs --include /blog --max-pages 80
npx llms-txt-doctor generate example.com --out - > llms.txt   # stdout

Option	Default
--out <file>	llms.txt	- for stdout
--max-pages <n>	50	page cap
--concurrency <n>	5	parallel fetches
--include <prefix>	—	repeatable path allowlist
--exclude <prefix>	—	repeatable path blocklist
--all	off	keep auth/pagination pages that are skipped by default
--title / --description	auto	override header lines

Discovery order: robots.txt Sitemap: lines → /sitemap.xml / /sitemap_index.xml (indexes followed) → homepage links. Same-origin only.

Output shaping (all automatic): auth flows and /page/N pagination are skipped (--all keeps them), legal boilerplate lands in the spec's ## Optional section, and localized paths like /zh/... collapse into one native-named section per language (e.g. ## 中文 (Chinese)), sorted after the default-language sections.

check <url>

npx llms-txt-doctor check example.com
npx llms-txt-doctor check example.com --json   # for scripts/CI

Exit code 1 if any check fails — usable as a CI gate.

llms.txt check — https://example.com/

  ✓ /llms.txt found (text/plain, 4.2 KB)
  ✓ H1 title: "Example"
  ⚠ Missing `> summary` blockquote after the title
  ✓ 3 section(s), 42 link(s)
  ✓ Links alive: 20/20 checked
  ℹ Covers 42 of 180 sitemap URLs (23%)
  ⚠ Site has 60 Chinese page(s) in its sitemap but llms.txt lists none of them
  ℹ llms-full.txt: not found (optional)

  0 failed, 2 warning(s)

Programmatic API

import { generateLlmsTxt, checkSite } from 'llms-txt-doctor';

const { content } = await generateLlmsTxt({ url: 'https://example.com' });
const report = await checkSite({ url: 'https://example.com' });

Notes

• Node 18+. Zero runtime dependencies.
• The generator never invents content: titles and descriptions come from the pages' own <title> / og: / meta tags.
• Prior-art survey that shaped this tool: docs/prior-art.md.

Who's behind this

Built by SeenForAI — we track how AI models (ChatGPT, Claude, Gemini, Perplexity, Doubao, Kimi, DeepSeek) describe brands. Checking what AI actually sees is our whole thing; this is the open-source corner of it. If you want to see how those models answer questions about your brand, the free checker runs without signup.

License

MIT