Draft an llms.txt from your own pages, without inventing anything about you
A prompt that turns your page list and real figures into a valid llms.txt, marks every gap instead of filling it, and says which pages do not belong in it.
- Works in
- Claude
- You need
- Your page titles and URLs, or a sitemap · What you sell, in one sentence · Your real prices, limits and dates
- Written for
- llms txt generator
Scored by our own engine
This page, run through the audit we sell. Measured 11 August 2026.
An llms.txt is a short markdown file at the root of your domain, listing what you are and which of your pages are worth reading. Writing one is twenty minutes of arrangement and one hour of arguing about which pages matter. This prompt does the arrangement, marks every fact it does not have instead of inventing one, and hands back the pages it decided to cut.
The three lists at the end are the point
The file itself is a formatting job. The lists underneath it are the analysis.
CUT shows you which of your pages a reader would not need before buying. It is an uncomfortable list and it is usually right.
MISSING names the questions nothing on your site answers. This is a content gap analysis derived from your own structure rather than from a keyword tool, and on most sites it turns up the same two or three obvious absences: what the thing costs in a form somebody can quote, what it does not do, and who it is not for.
RISKY lists every line that becomes wrong when a price changes. That list is the argument for generating the file from the same source your pricing page renders from, which is how ours is built and the single decision that keeps it honest between deploys.
Why it will not fill a gap for you
The prompt is instructed to write [NEEDS FACT: ...] rather than estimate. This is stricter than it sounds and it is deliberate.
Every other document on your site is read by a person who brings judgement. This one is written to be read by a machine at the exact moment it decides how to describe what you sell and at what price. A number invented here is not a small error in a paragraph, it is a small error positioned to be quoted with your name on it.
The same rule appears in our system prompt template for the same reason. A marked blank is fixed in seconds. A plausible invention is fixed after a customer quotes it back to you.
What the file cannot do
It does not grant access. If robots.txt disallows GPTBot or ClaudeBot, publishing an llms.txt achieves nothing, because the crawler that would read it is not allowed to fetch it.
That combination is not rare. In our study of AI crawler access across 1,048 of the most visited websites, 121 sites published a real llms.txt and 11 of them were blocking at least one major AI crawler in robots.txt at the same time. The file and the block were almost certainly added by different people, months apart, neither knowing about the other.
Check both in the same sitting. Our llms.txt checker fetches the live file and reads the response rather than trusting the status code, which matters more than it should: in that same crawl, 162 domains returned HTTP 200 for /llms.txt while serving an HTML page rather than a file. A checker that counts status codes would have reported more than double the real adoption rate. The AI crawler checker covers the other half.
For the argument about whether any of this is worth doing at all, including the honest state of the evidence, see llms.txt: what it is and whether it works.
You are drafting an llms.txt file: a markdown document served at the root of
a domain that tells an AI system what this site is and which pages are worth
reading. I will give you a page list and a set of facts. Use only those.
Produce the file in one code block, in this structure:
# [Site or company name]
> One or two sentences saying what this is and what it sells. Put the two or
> three hardest facts, the ones with numbers in them, inside this blockquote.
> If an assistant quotes nothing else from this file, this is what it quotes,
> so it must be accurate and complete on its own.
## [Section name]
- [Page title](url): one sentence saying what a reader gets from this page,
written so it is useful to somebody deciding whether to open it.
## Facts worth quoting
- Flat declarative statements, one per line, each containing the figure and
the thing the figure describes, in the same line.
## Notes for AI systems
- The date these facts were checked, and one line naming the page that is
authoritative if this file and the site ever disagree.
Rules for building it:
1. Select, do not list. Choose the 15 to 30 pages that answer what somebody
would want to know before buying or using this. A file that links to
everything has made no editorial decision and is worth no more than a
sitemap. If I give you more pages than that, cut, and tell me underneath
the code block which ones you cut and why.
2. Group the selected pages into 3 to 6 sections named after what a reader
wants, not after our internal structure. Products, pricing, how it works,
free tools, documentation, research. Never a section called Other.
3. Every figure must come from the facts I gave you, with its unit. Never
round one, never average two, never carry a figure from one product to
another.
4. If a section needs a fact I did not give you, write
[NEEDS FACT: what is missing] on that line and keep going. Do not
estimate, and do not write around the gap so it disappears.
5. Never write marketing adjectives. Leading, innovative, seamless,
world class, best in class. This file is read by a machine deciding how
to describe us, and an adjective displaces a fact that would have been
quoted.
6. Keep each description to one sentence. The file should be under 2,000
words unless the page list genuinely justifies more.
After the code block, give me three short lists:
- CUT: pages I gave you that are not in the file, with a reason of five
words or fewer each.
- MISSING: things a reader or an assistant would obviously want that no page
of mine appears to cover. This is a content gap list and it is the most
useful output here.
- RISKY: any line in the file that would become wrong if a price, a limit or
a date changed, so I know what to regenerate this from rather than
maintain by hand.
Do not ask me anything first. State assumptions, labelled, and continue.
What we sell, in one sentence: [ONE SENTENCE]
Facts, each with its unit and the date checked: [PRICES, LIMITS, NUMBERS]
Pages: [PASTE TITLES AND URLS, OR A SITEMAP]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE SENTENCE]- What you sell, written the way you would want it quoted back to you, because the blockquote at the top of the file is the passage most likely to be lifted whole. If you sell two different things, say both here rather than picking one, or the file will describe half your business.
[PRICES, LIMITS, NUMBERS]- Copied from wherever they are authoritative, usually your pricing page, each with its unit and the date you checked. This is the only place figures may come from. A file built from remembered prices is a confidently wrong answer waiting to be quoted at a customer.
[PASTE TITLES AND URLS, OR A SITEMAP]- Titles and URLs, or the raw sitemap. Give it more than you want in the file and let rule 1 cut: the CUT list it returns is a useful second opinion on which of your pages actually earn their place.
How to run it
- 01Collect the facts before the pages
Open your pricing page and copy the real numbers with their units. Everything else in this process is arrangement. The figures are the part that can be wrong in a way that costs you a customer, and they are the part a model will happily supply if you leave them out.
- 02Give it more pages than you need
Paste the sitemap. The selection is the valuable work, and watching which pages get cut tells you something about your own site that a full list never would.
- 03Read the MISSING list first
The gaps it names are usually more useful than the file. They are the questions a reader would ask that nothing you publish answers, which is a content plan derived from your own structure rather than from a keyword tool.
- 04Fill the NEEDS FACT markers by hand
Every marker is a real fact you have not written down anywhere machine readable. Fill them from the source, not from memory, and if you cannot fill one, delete the line rather than softening it into a claim.
- 05Generate it from your source of truth, then check it live
The RISKY list names every line that goes stale when a price moves. Wire those to the same source your pricing page renders from if you can. Then publish at /llms.txt and confirm what a fetcher actually receives, because a server that answers every unknown path with your homepage will look like it has the file when it does not.
Questions people ask
What is llms.txt?
A markdown file served at the root of a domain, at /llms.txt, listing the pages worth reading and stating plainly what the site is. It was proposed in September 2024 and it is a convention rather than a ratified standard. It controls nothing: robots.txt decides access, and this file only offers editorial guidance to whoever chooses to read it.
Do assistants actually read llms.txt?
No major AI operator has published documentation confirming their assistant fetches it during retrieval, and Google has said explicitly that no AI text file is required for its AI features. The argument for publishing one is the cost, which is close to zero, rather than the evidence, which does not yet exist. We wrote the long version of that argument up separately, and this prompt exists because the file is cheap, not because it is proven.
Why does the prompt refuse to fill in missing figures?
Because this file exists to be quoted verbatim. A wrong price in an ordinary paragraph is a mistake. A wrong price in the one document you published for machines to read is a mistake positioned exactly where it will be repeated, and it will be repeated with your name attached, which is why the NEEDS FACT marker is required instead.
How many pages should be in it?
Between fifteen and thirty for most businesses. The value of the file is the editorial decision, so a list of four hundred URLs is worth no more than the sitemap you already publish. Twenty links each with a sentence of context is more useful to a machine than every page you have.
Should I publish llms.txt before fixing anything else?
No. If your robots.txt blocks the AI crawlers, or your pages only exist after JavaScript runs, this file is decoration on a locked door. Check those two things first. In our own crawl of 1,048 of the most visited websites, 11 of the 121 sites publishing an llms.txt were simultaneously blocking at least one major AI crawler in robots.txt, which is the same contradiction at scale.
Unsubscribe in one click. We never pass your address on.