- Home
- Skills
- Writing & Content
- Source Quote Checker
Works with the AI Agents you already use
Source Quote Checker
Give every claim in a draft its source URL and exact quote; a script then checks each quote is still on the page: FOUND, NOT FOUND, SITE DOWN, NO SOURCE.
$5
Source Quote Checker
Example session with this skill installed
The paragraph and the user are invented. The two web pages are real, read by the script on 24 October 2026.
User message:
Can you check the claims in this paragraph before I publish it on our agency blog? Today is 24 October 2026.
"A good website is built on standards, not guesses. WCAG 2.1, the W3C standard for accessible web content, is organised around four principles and 12 guidelines. If your site has more than 50,000 pages, one sitemap will not be enough: a single sitemap file may hold at most 50,000 URLs and 50MB. And remember that most visitors leave a page within 15 seconds, so put the important thing first."
The two pages I used: https://www.w3.org/TR/WCAG21/ and https://www.sitemaps.org/protocol.html
- Read your context and instructions
- Compiled the source quote checker
Source check: agency blog paragraph (WCAG and sitemaps) · 24 October 2026
Result: 7 claims · 6 FOUND · 0 NOT FOUND · 1 NO SOURCE · 0 SITE DOWN · 0 HAND-CHECKED
Claims file: claims.json (script run: yes)
| # | Claim | Status | Source | What to do |
|---|---|---|---|---|
| 7 | Most visitors leave a page within 15 seconds | NO SOURCE | none found | The pages you gave say nothing about this. Add a primary source (a study or dataset that states this figure, with its date and what it measured), or soften the sentence, or cut it |
| 3 | WCAG 2.1 has 13 guidelines (your draft said 12) | FOUND | https://www.w3.org/TR/WCAG21/ | Your sentence is wrong: fix "12 guidelines" to "13 guidelines" |
| 1 | WCAG 2.1 is a W3C Recommendation for accessible web content | FOUND | https://www.w3.org/TR/WCAG21/ | The page says "Recommendation", not "standard". Consider "W3C Recommendation" in the text, or keep "standard" knowing the page does not use that word |
| 6 | A site listing more than 50,000 URLs needs more than one sitemap file | FOUND | https://www.sitemaps.org/protocol.html | Your draft says "50,000 pages"; the page counts URLs listed in the sitemap. Say "more than 50,000 URLs" or check that your pages map one to one to URLs |
| 2 | WCAG 2.1 is organised around four principles | FOUND | https://www.w3.org/TR/WCAG21/ | Nothing |
| 4 | A single sitemap file may hold at most 50,000 URLs | FOUND | https://www.sitemaps.org/protocol.html | Nothing |
| 5 | A single sitemap file may be at most 50MB | FOUND | https://www.sitemaps.org/protocol.html | Nothing. The page adds that the limit applies to the file once uncompressed (52,428,800 bytes) |
Sentence fixes (the rewrite for each NOT FOUND, NO SOURCE or scope problem, with the quote it rests on)
- "WCAG 2.1 ... is organised around four principles and 12 guidelines" becomes "four principles and 13 guidelines". Quote: "The 13 guidelines provide the basic goals". The page read is the W3C Recommendation of 06 May 2025 (the latest published version of WCAG 2.1 at the time of the check).
- "the W3C standard for accessible web content": the page calls it a "W3C Recommendation". Suggested: "the W3C Recommendation for accessible web content".
- "If your site has more than 50,000 pages, one sitemap will not be enough": the quote is "If you want to list more than 50,000 URLs, you must create multiple Sitemap files". Suggested: "If your sitemap needs to list more than 50,000 URLs, one sitemap file will not be enough".
- "most visitors leave a page within 15 seconds": no primary source, so no rewrite based on a quote. Either find the source or write something you can support, for example "many visitors leave quickly" with a source, or cut the sentence and keep the advice "put the important thing first".
Not checked: opinions, predictions and the author's own experience: "A good website is built on standards, not guesses", "put the important thing first". Scripts cannot check whether the sources are right or whether the sitemap limits suit your own site; the script only confirms that the words are on the pages. The guideline count and the sitemap limits are technical rules, so a person who knows web accessibility and SEO should read the two sources too.
Connects securely to your tools. The creator never sees your data.
About this skill
Before you publish a post, landing page, report or newsletter, give each factual claim its primary source and an exact quote, then check that the quote is still on the page. Claude does the claim-by-claim work. A short Python script does the checking, so FOUND means the words are on the page, not that a model believed so.
What you get
- Every factual sentence of your draft listed as a claim, with a source URL and the exact words from that page, saved in a claims.json file you can run again whenever you like
- A script (Python 3.9+, no packages to install; PDF pages need pypdf) that downloads each page once and reports FOUND, NOT FOUND, SITE DOWN, NO SOURCE or HAND-CHECKED for every quote
- Claims with no primary source flagged, and links to encyclopedias, forums or AI chat shares flagged as weak sources
- For a NOT FOUND, the closest text on the page, so you can tell a page that changed from a quote that was wrong
- A --find option that prints the page's own sentences around a word, so you can copy an exact quote instead of retyping one
- A report with a sentence fix for each problem. Claude reads each sentence against its quote for scope, "may" against "must", and dates
How it works
- Claude lists the claims in your draft. If your setup gives it web search or fetch, it finds a primary source for each (the law, the regulator, the company's own documentation, the study); otherwise you give it the URLs, and a claim with no source stays NO SOURCE.
- It copies the shortest exact quote from the page and writes the claims file.
- You (or Claude) run: python scripts/check_claims.py claims.json
- You get a table, one row per claim, and fixes for every NOT FOUND, NO SOURCE and over-claimed sentence.
A method we use ourselves
We check the legal and price claims on our own website with the same claims-file method. This skill is that method, made general: no site-specific rules, no files from our project.
Tested before listing
- The script has 41 automated checks run against a local test server: exact match, curly quotes and non-breaking spaces, text hidden in scripts and comments, a changed number, an empty source, maintenance pages, pages that refuse scripts, unreachable hosts, PDFs, other character sets, redirects, slow servers, exit codes
- The skill was run on three examples with real pages (W3C WCAG 2.1, the sitemaps.org protocol, RFC 2119): a draft with one wrong number and one claim without a source, a claims file with one quote gone from its page, one sentence that says more than its quote and addresses that never open, and a draft with no sources at all. The three runs are in examples/
What it will not do
- It proves that words are on a page. It does not prove the source is right, current or the best one
- It does not check opinions, predictions or your own experience
- Pages that need a browser to show their text usually come back as SITE DOWN (a page that shows a little text without JavaScript can come back NOT FOUND: open it in a browser before you correct anything). Sites that refuse scripts, and pages over 25 MB, come back as SITE DOWN too: read them in a browser and mark the entry "checked"
- It is not a legal, medical or financial review: for those, a person who knows the field should read the source too
Good to know
- For Claude Code and the Claude app with code execution. The script needs network access to the pages you list and requests only the addresses in your claims file (and where they redirect, http and https only); without code execution the skill checks each page by hand and says so
- Claude reads your draft like any other message and the script sends nothing but page requests
- By Suivel (suivel.fr)
Checks that an exact quote is on a page, after ignoring case, spacing and typographic quotes. It does not judge whether the source is right, recent or the best one, and it cannot see that a sentence goes further than its quote: the skill asks Claude to read each sentence against its quote, and you should do the same for anything important. Pages that need JavaScript usually come back as SITE DOWN, but a page that shows a little text without it can come back NOT FOUND: open it in a browser before you correct anything. Sites that refuse scripts, pages over 25 MB, scanned PDFs and PDFs without the pypdf package come back as SITE DOWN; read them in a browser and add a checked date. Redirects are followed (http and https only), so a page can end up somewhere other than the address you wrote. Opinions, predictions and your own experience are not checked. Legal, medical and financial statements need a human reader who knows the field.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 24 days ago
Requires
Permissions
File Scopes
The script requests the addresses written in the claims file (and where they redirect, http and https only) and sends nothing else. Terminal / Shell runs the script, Read Files reads claims.json, Write Files lets Claude write claims.json and the report, Network Access downloads the pages you list.