Schema Sheriff: structured data checked against what your visitors actually see
Schema Sheriff is a skill for Claude that audits, fixes and generates structured data (JSON-LD, schema.org) and llms.txt. It extracts every JSON-LD block from your pages, validates it, then compares what the markup claims with what the page shows: FAQ text, prices, ratings. It re-reads Google's current policy before saying anything about rich results.
One of the 14 skills in the pack
Ask it in your own words
- "Check my schema markup."
- "Should I add llms.txt?"
- "my stars disappeared from google, did i break something"
- "Write the Organization and WebSite JSON-LD for my home page."
What happens next
- It re-reads the policy. Before stating what Google does with a type of markup, Claude opens the official page again and dates what it read.
- It extracts and compares. jsonld_check.py, a read-only checker, pulls every JSON-LD block from a sample of pages. Each finding is then confirmed by hand on the visible text, both ways.
- It audits llms.txt. llms_check.py checks structure, links and wording. A missing llms.txt is not counted as a defect.
What you get back
- A five-line verdict, ending on the first action
- A dated policy table, with the source of each line
- A findings table: severity, page, literal evidence, fix, effort
- The facts that disagree between page, markup and llms.txt
- A fix plan ordered by risk, applied where the markup is generated, and what you must run in a browser
Run on this very site
The structured data of the home page, compared with what the page's HTML holds.
warning | FAQ_ANSWER_MISMATCH | the question is on the page but the visible answer shares only 0% of the marked-up answer's phrases [...]
Faut-il un outil SEO payant ? | exact | 0% | answer wording differs
after the fix: Faut-il un outil SEO payant ? | exact | 100% | matchThe folded answers of the FAQ were not in the page's HTML at all: a crawler could not read them. They are now, and all seven match.
What it will not do
- It reads static HTML: markup, reviews or counters that JavaScript writes into the page are not seen.
- It cannot tell whether reviews are genuine, only whether the figures are visible. It asks where they come from.
- It will not write ratings, FAQ answers or awards the page does not show, not even as an example.
- Valid markup promises no rich result, ranking or AI citation.
What it needs
Claude Code, Claude.ai, Claude Desktop or Cowork. The two checkers need code execution with Python 3.9+, and web access unless you point them at a local build. Without code, Claude reads the raw page source by hand.
How to install a skill in Claude
Guides
Where it hands over
- Name Drop: To test whether AI assistants mention your brand, which this skill does not measure. You come back here for the fixes.
- Site X-Ray: For the rest of the site (crawlability, titles, links): this skill is not a ranking audit.
Questions about this skill
Should I add llms.txt?
It answers that first, in five dated lines: optional, unproven, ignored by Google Search, and worth publishing only if someone will keep it true.
How many pages does it look at?
A sample: three to six representative pages you name, or pages spread evenly over the sitemap, capped at 25 by default. The report opens with what was not audited.
It comes with the 13 others.
One pack, one payment: the 14 skills, their scripts, templates and reference notes.
One payment, no subscription. In your inbox within a minute, as a rule.