An open, free educational protocol for questioning claims you see online - using AI tools you already have.
Enter what you heard, pick three tools, and we’ll walk you through validator → referee → meta-referee. Not for medical, financial, or legal decisions.
CrossCheck is a simple process: take anything you want to verify (a headline, rumor, quote, question, or screenshot text), run it through a short chain of independent AI chats, and still verify the result against real primary sources yourself. It takes about ten minutes.
When your CrossCheck record is complete, you can send your claim and results to the moderated community feed. Someone else working a similar topic can see your wording, pasted replies, and how far you took the workflow — so they are not reinventing the same run from scratch. Nothing in the feed is “final”: if you disagree, you are welcome to run your own CrossCheck (including from a Shared run card) or flag a post for moderator review using the controls on the results page.
The prompts and paste fields below are for educational use. Read the legal documents (open in a new tab if you prefer), then confirm acceptance. The form then opens in three short steps: define your claim, pick your three-tool lineup, then run the CrossCheck prompts.
Tools locked
Check the box under "Before you begin" to accept the terms. You will then unlock the form in short steps (claim, lineup, CrossCheck prompts).
How sections unlock: Later panels stay collapsed with a one-line note until the prior step is done. CrossCheck uses tight prompts (short answer + small summary block, then paste, referee, meta-referee). Legal/resources stay in collapsed panels at the bottom.
Step 1 — Define your claim
Paste or type the exact wording you want to verify — a question, statement, transcript snippet, political line, rumor, or anything else. One box only. Leave it empty to keep the built-in Lego safety example for every prompt below.
The claim field is empty — prompts below use the Lego sample until you type your own wording.
Type your claim text above, or clear the field again to return to the worked example.
Your claim text is set — prompts below use your wording.
You can continue with the built-in example (empty field) or type your own claim — then click Continue to lineup.
Step 2 — Lineup is collapsed until you click Continue to lineup under Step 1.
Step 2 — Pick your AI tools
Pick three AI tools — the steps below will update with your selections. Free options are listed first. Prompts ask models for concise, structured answers; referee and meta-referee copy includes short inline rules unless you pick Perplexity or Copilot (often web-capable — then we mostly link AGENTS.md).
Confirm or change each slot below. Every prompt later will name the tools you pick here.
No problem — you can open separate new chats in the same AI for each step. The key is that each conversation starts fresh, so no answer is influenced by the previous one. Even one tool used in three temporary tabs will work.
One AI, several tabs. When you only have one provider, open three temporary chat tabs and run validator, referee, and meta-referee in separate fresh chats so later steps are not primed by earlier ones in the same thread.
How to line them up.
Step 3 — CrossCheck is collapsed until you click Continue to CrossCheck steps under the lineup.
Step 3 — Run the CrossCheck
Open a new chat. The model checks the single claim you defined above against primary sources (agencies, studies, official labeling — not random blogs). Same one-shot Part 1 + summary block pattern.
Copy the prompt below into a new chat in ChatGPT. When the model finishes, paste its entire reply into the box under this step.
Use ChatGPT in a fresh chat only — you’ll paste its full message below.
Paste the primary validator’s full reply
Include the short CrossCheck summary block at the end. Step 2 stays collapsed until this box has text.
Paste the model’s full reply from Step 1 here.
Validator reply captured — Step 2 unlocks.
Step 2 — Referee is collapsed until the primary validator paste box above is filled.
Open a new chat in your referee tool. The bundled prompt includes the claim and the pasted validator output.
Paste the Step 1 reply above first — otherwise the referee prompt still contains placeholders.
Copy the prompt into a new chat in Grok, then paste the referee’s full reply below.
Open Grok. Ask for the structured sections in the prompt — this pass judges how well the validator supported the claim.
Paste the referee’s full report
The meta-referee prompt in Step 3 will pull from this box automatically when you copy.
Paste the referee’s full reply from Step 2 here.
Referee report pasted — Step 3 unlocks.
Step 3 — Meta-referee is collapsed until the referee report is pasted.
New chat in your third tool. The prompt embeds the claim, the validator paste, and the referee report so the model can stress-test the middle layer.
Paste the referee report in the box above so the prompt below is complete.
Copy the prompt into a new chat in Claude — then open its links yourself.
Take everything to Claude. This is the answer you trust most — after you click through to the sources it cites.
Paste the meta-referee’s full reply
Optional for export, but strongly recommended: this is the second independent check on the referee layer.
You’ve pasted the primary validator and the referee — enough to save or share a structured record. Add the meta-referee when you can; it completes the chain. Use JSON for tooling; Markdown or plain text for notebooks or email.
Sharing & repeating checks: Publishing a record can spare others from asking the same question blindly — but running CrossCheck again later still matters: models drift, sources update, and a second pass is legitimate verification (same way scientists replicate).
Optional step: Paste your Step 3 meta-referee response above to include that layer in exports — until then, that section stays empty in the files below.
Publishing saves your record as pending. Shared runs lists approved posts only — submissions are moderated and do not appear there until someone approves them. See CONFIGURATION.md for moderation and api/moderate.php. Solo dev boxes may set CROSSCHECK_AUTO_APPROVE=1 instead.
Three linked reviews. One claim. You didn’t stop at the first confident answer — you forced a referee, then a meta-referee, and you still owe the primary sources a click.
For claims that touch health, safety, or money, you can still use CrossCheck as a literacy exercise - but decisions about treatment, diagnosis, medications, diet, investments, taxes, or contracts belong with qualified professionals, not with this page or any AI. For extra scrutiny on general claims, take the final write-up back to ChatGPT, Grok, and other tools and ask each to find errors. The more independent eyes, the better.
No matter how many AIs agree, click through to the actual primary source — the manufacturer's website, the government agency, the published study — and read it yourself. AI consensus is a useful guide. Primary sources are the answer.
Canonical legal text: AGREEMENTS.md. License for materials in this repo: LICENSE (CC BY 4.0).
CrossCheck (this website, the interactive prompts, and related materials) is offered for educational and academic purposes only: to help you practice questioning information you encounter online and to encourage you to consult original sources yourself.
Not professional advice. Nothing here is medical advice, mental-health counseling, diagnosis, or treatment; nothing here is financial, investment, tax, accounting, insurance, or legal advice. CrossCheck does not verify claims to a professional or regulatory standard. Do not use this site or any AI output in place of a qualified clinician, attorney, accountant, financial advisor, or other licensed expert. If you think you may have a medical emergency, call your local emergency number.
No warranty. The site and protocol are provided "as is" without warranty of any kind, express or implied. We do not guarantee accuracy, completeness, fitness for a particular purpose, or that AI tools will produce correct citations or conclusions.
Acceptance. By checking the "Before you begin" box above you confirm you have read AGREEMENTS.md and the license and accept them. This section is a summary; the Markdown files control if they differ. If you do not agree, leave the box unchecked and do not use the tools below.
Educational use only. Terms & agreements · On-page summary. Not medical, financial, or legal advice.