<aside> 🔎

What you're making: a research answer where every claim comes with a link you can open, the exact sentence on that page that backs it, and a second check by an AI that never saw the draft.

Why this exists: AI search repeats bad sources with a straight face. Google traced its AI Overview errors to satire, joke forum posts and topics with almost no good pages. A 2026 audit of 55,393 Google searches found 11% of AI Overview claims weren't backed by the page they cited. No single magic phrase fixes this. A setup plus a checking routine does, and every step below comes from a peer-reviewed paper, official Anthropic, OpenAI or Google documentation, or an independent benchmark.

What you need:

The order never changes: clean the question, research, split into claims, verify in a fresh chat, check the links, revise. The checking never happens in the chat that wrote the draft.

</aside>

<aside> 🔎

The Short Version

  1. Clean the question. Strip your opinion out and split it into facts that can be searched one by one.
  2. Research with rules. New chat, thinking model, search on, master prompt. Every claim gets a URL, a quote and a source tier.
  3. Split into claims. New chat. One checkable fact per line.
  4. Verify in a fresh chat. The other company's model, search on, 10 claims at a time, told to try to prove each one wrong.
  5. Check the links. Every URL loads, every quote is really on the page, every paper exists.
  6. Revise. Keep only what passed. Add nothing new.
  7. Review by hand. Numbers, quotes, medical, legal and money facts, niche topics. </aside>

🧭 Step 1 · Clean up the question

🔍 Step 2 · Research with evidence rules

🧩 Step 3 · Split the draft into single claims

⚖️ Step 4 · Verify each claim in a fresh chat

🔗 Step 5 · Check links, quotes and papers

✍️ Step 6 · Revise without adding facts

👀 Step 7 · Review by hand where AI checkers miss

🗞️ Bonus · The Prompt Pack

📚 More context