Build Your Own Research Team

Connect: LinkedIn (linkedin.com/in/lsartain) · Workshop sign-up

I built claim-check as a small research database. Claude pulls facts from saved public sources, and a plain Python script checks each one against the source. The project is at github.com/lsarta/claim-check.

What you need

• A GitHub account.
• A paid Claude plan (Pro or higher).

Set up (5 steps)

1. Create a GitHub account at github.com. You'll slide a puzzle to prove you're human, then confirm your email. This takes 5-10 minutes.
2. Get Claude Pro or higher at claude.ai (the free plan has no Claude Code).
3. On github.com/lsarta/claim-check, click the green Use this template ▾, then choose Create a new repository (not Open in a codespace).
4. Open claude.ai/code, connect GitHub (approve the screens), and pick your copy.
5. Type this:

Run the checker and show me the report.

You should see 8 claims: 7 verified, 1 held.

Prompts from the session

Run the checker and show me the report.

You should see 8 claims: 7 verified, 1 held.

Run a tamper test: in the trade capture, change the word "deficit" to "surplus" once, run the checker, and show me the report. Then put the file back and run it again.

You should see c6, c7 and c8 held (fingerprint mismatch), then back to 7 verified, 1 held.

I want to research: [your question].

Here is text I copied from [page title], at [page address]. Save it as a new capture, pull out the five numbers that best answer my question with their exact quotes, run the checker, and show me what was verified, what was held, and why.

[paste the text here]

You should see the five numbers, each marked verified or held, with the reason.

Undo your changes and run the checker again.

You should see 8 claims: 7 verified, 1 held.

Try your own source

What makes a good source:

• An official release that stays the same after publication.
• Numbers written out in sentences, since the checker needs a quote.
• A public page that does not require a login.
• The agency's own release rather than news coverage of it.

BLS, BEA, Census and the Federal Reserve all work well. For example, try the Census new residential construction release (PDF).

Keep it current

A routine is a scheduled run that checks the pages you choose for new releases, saves a copy of each one, adds claims and runs the checker, then leaves the results for you to review. To set one up, follow ROUTINE.md in the repo.

Connect: LinkedIn (linkedin.com/in/lsartain) · Workshop sign-up