# Your llms.txt matches what your site actually ships

> Your llms.txt matches what the site ships. Checks it, reports what is missing.

Loop: `llms-txt-build`
Category: AI visibility
Autonomy: auto
License: MIT
Source: https://vibemarketers.ai/loops/llms-txt-build

Checks the published llms.txt against the real page list and reports the published pages an agent cannot discover from it.

## What you need

- none required

## Goal prompt

Give this to your agent. Replace anything in <angle brackets> with your own.

```
Read the published llms.txt for <site> and the list of pages the site
actually ships. Report every published page that llms.txt does not list, and
every URL llms.txt lists that does not exist. Check whether it points at the
machine-readable surfaces: raw files, feeds, and any API. Write the report
with a date.
```

## Verifier prompt

Run this as a second, separate agent. It grades the first one's work and must be
able to fail. A checker that never fails is not a checker.

```
Open the report. FAIL if a page that ships is reported as listed, if a listed
URL that 404s is reported as fine, if the report omits any published page, or
if it has no date. Otherwise PASS.
```

## Stopping condition

Every published page has a listed or missing verdict in the report, and the verifier marks PASS.

## How to run it

1. Save this file into your project, for example `.claude/loops/llms-txt-build.md`.
2. Open your agent in that project (Claude Code, Cursor, or Codex).
3. Paste the goal prompt, with your own values in place of the angle brackets.
4. When it stops, start a fresh session and paste the verifier prompt.
5. Read the verdict before you accept any change.

## What happened when we ran it

Last run 2026-09-25 on vibemarketers.ai.

- Baseline: An llms.txt written by hand at launch, listing 31 URLs, with no check that it matched what the site had since shipped.
- Result: 31 URLs listed against 32 pages. Every published loop and every term page was listed. 3 key pages were missing: the catalog itself, the runs page and the questions page. All 3 were added and a rerun found 0 missing.
- What this does not prove: Being listed in llms.txt is not evidence that any agent reads it or acts on it. This checks coverage, not effect.
- Cost: 0 human touches, 4 minutes, $0.00
- Raw log: https://github.com/genwfurukawa/vibemarketers-v2/blob/main/lab/runs/2026-09-25-llms-txt-build.json

## Inside a company

- Problem it solves: llms.txt is written once and then drifts, so agents get an out of date map of the site.
- Source of truth it reads: The published llms.txt and the real page list from the build.
- What it may do: Read llms.txt and the page list, compare, write a report. Never publish a rewritten file without approval.
- Who approves: The owner approves any change to llms.txt before it ships.
- Who owns it: Technical SEO owner
- How you measure it: Share of published pages discoverable from llms.txt.

MIT licensed. Copy it, change it, run it. No attribution required.