Is llms-full.txt part of the llms.txt specification?

Not in the text. The v2 proposal at llmstxt.org, last modified August 10, 2026, defines the /llms.txt file, markdown twins at the same URL with a .md extension, and the rel=alternate and rel=describedby link relations. A search of that page for the string llms-full returns nothing. llms-full.txt is a widely adopted convention — a concatenation of your markdown pages into one file — but a client that only implements the written spec has no obligation to look for it, which is why the index file has to link it explicitly.

How many tokens is a large llms-full.txt file?

Divide the byte count by roughly four for an English-markdown character-based estimate. On that basis the 67,649,033-byte file at developers.cloudflare.com is in the region of 16.9 million tokens and the 43,107,672-byte file at platform.claude.com is around 10.8 million. Run your own tokenizer if the number matters for billing or retrieval planning — the estimate is a planning figure, not a measurement of what a specific model will count.

Should a small site publish an llms-full.txt at all?

Usually not. Below roughly fifty pages the curated index is the whole story, and a full file duplicates content an agent can fetch page by page from the links it already has. Publish one when the site is documentation-shaped — hundreds of pages that agents need in combination — and even then split it per section rather than shipping one file, so a reader can pull only the part it needs.

Format & Spec · October 2026

llms-full.txt Size and Structure: The Full File Nobody Specified

Everyone references llms-full.txt. The specification does not. We fetched the file from eight live documentation sites on October 11, 2026 and measured every byte: the largest is 67.6 MB before compression, the smallest still 433 KB, and four of the domains that serve one never mention it in their own /llms.txt.

The Spec Never Defined llms-full.txt

The v2 proposal at llmstxt.org, last modified August 10, 2026, describes the /llms.txt file, markdown twins served at the page URL with a .md extension, and two link relations — rel="alternate" type="text/markdown" for the twin and rel="describedby" for the index that covers a page. We searched the page for the string llms-full on October 11, 2026: zero matches.

That absence is deliberate. The spec states the index "stays small enough to fit in context. The detail lives behind the links, and is fetched only when needed." A single file holding every page of a large docs site is the opposite design, and it survives as a convention rather than a requirement — which is exactly why a client that implements only the written spec will never look for one unless your index points at it. The comparison between the curated index and the full dump is in llms.txt vs llms-full.txt.

What Eight Live Full Files Actually Weigh

Every request followed redirects and timed out at 30 seconds. Sizes are uncompressed bytes on the wire after decoding:

Host and path	Bytes	gzip transfer	Structure signal
developers.cloudflare.com/llms-full.txt	67,649,033	11,527,423	29,052 H2 headings; per-section files exist
platform.claude.com/docs/llms-full.txt (301 from docs.anthropic.com)	43,107,672	—	starts with a "Root URL" page, then full rendered docs
nuxt.com/llms-full.txt	4,893,417	1,291,540	3,586 H2 headings
ai-sdk.dev/llms-full.txt	4,724,729	—	byte-identical to /llms.txt on the same host
vitest.dev/llms-full.txt	1,382,071	371,248	215 `url: /` page markers
svelte.dev/llms-full.txt	1,261,050	357,726	220 H1 page headings under a SYSTEM preamble
vuejs.org/llms-full.txt	953,556	314,771	94 `url: /` page markers
vite.dev/llms-full.txt	432,841	127,916	42 `url: /` page markers

Four hosts we probed do not serve one at all: docs.stripe.com returned 404 with 23,881 bytes of HTML, www.anthropic.com returned 404 with 20,584 bytes of an error page, openai.com returned 403 with 9,698 bytes, and docs.astro.build answered 404 with no content type and a zero-byte body.

Compression Buys Four to Six Times, Not Enough

All six hosts that served a real file sent content-encoding: gzip when we offered it. The saving is real and consistent — Cloudflare's 67.6 MB arrives as 11.5 MB, Nuxt's 4.9 MB as 1.29 MB, Vite's 433 KB as 128 KB. On a character-based estimate of bytes divided by four, that still leaves the two largest files at roughly 16.9 million and 10.8 million tokens respectively, with Nuxt and the AI SDK near 1.2 million each.

Treat that arithmetic as planning, not measurement: it is a byte-to-token approximation, not a tokenizer run, and the ratio shifts with code blocks and non-English text. The direction is never in doubt. Any file in the seven-digit token range can only be read in slices, so the file has to be built to be sliced. The size-versus-verdict table for the index file itself is in llms.txt size limits.

Split It the Way Cloudflare Does

The 67.6 MB root file is not the only option that host publishes. Its product sections each carry their own full file, and all three we probed returned 200:

/workers/llms-full.txt — 5,127,319 bytes
/learning-paths/llms-full.txt — 1,841,842 bytes
/ai-gateway/llms-full.txt — 910,239 bytes

That is the pattern worth copying, and it is the one the spec's subpath rule already supports: an index at a path covers the URLs beneath it, and the most specific file wins. A reader that needs the gateway docs fetches 910 KB instead of 67 MB. Structure it as one file per product or per top-level section, keep the root full file optional, and give each section its own index so the split is discoverable. The same reasoning applies to a markdown-twin pipeline, which is where these files come from — see llms.txt for API documentation.

Only Two of Six Index Files Mention Their Own Full File

We pulled /llms.txt from each of the six hosts that serve a full file and counted references to it:

nuxt.com — three mentions, including a prose line: "Every documentation, blog and deploy page is also Markdown" followed by the full-file URL and a note that other versions are listed in full there.
svelte.dev — one entry, "Complete documentation", describing it as all Svelte and SvelteKit docs with examples.
developers.cloudflare.com, vuejs.org, vite.dev, vitest.dev — zero mentions of llms-full in the index file.

Nuxt's wording is the standard to aim for, because it tells a reader both what the file is and when to prefer it. And ai-sdk.dev shows the failure mode at the other extreme: its /llms.txt serves 4,724,729 bytes with the same MD5 (eebec45d272ac9891b3c5be10912c124) as /llms-full.txt on the same host, in text/markdown. The path meant for a curated index is a 1.2-million-token dump, so an agent that politely checks the conventional URL first pays the full price before it learns anything.

Three Commands Before You Publish One

The checks that matter are size, encoding and what the index actually links:

# 1. Size and content type of the full file
curl -sSL -o /dev/null -w '%{http_code} %{size_download} %{content_type}\n' \
  https://www.example.com/llms-full.txt

# 2. What gzip actually costs a reader (send Accept-Encoding explicitly)
curl -sSL -H 'Accept-Encoding: gzip' -o /dev/null \
  -w 'wire=%{size_download} encoding=%{content_type}\n' \
  https://www.example.com/llms-full.txt

# 3. Does the index point at it, and how many pages are inside?
curl -sSL https://www.example.com/llms.txt | grep -ci 'llms-full'
grep -c '^url: /' llms-full.txt

The third line is the one that finds most problems: either the index is silent and the full file is effectively invisible, or the count is in the thousands and the file should have been split. Build the index in the generator, verify paths and link targets in the checker, then measure the delivered bytes rather than trusting the file in your repository.

The verdict: llms-full.txt is a convention nobody wrote down, published at sizes that no context window holds, and referenced by only a third of the sites that pay to serve it. Ship it per section, describe it in one honest line in the index, and let the reader choose the 910 KB slice over the 67 MB whole.