Have several reviewers try to prove the work wrong
In an interview posted by Marina Mogilko, Felix Rieseberg, who helped build Claude Cowork at Anthropic, explained how he catches confident mistakes. He has Claude start five threads that each try to prove a document wrong, then a separate Claude judges which points hold up. Every number gets checked against a primary source, and he rarely lets a session grade its own work.
The lesson: Don't ask the same chat whether its work is right. Have separate reviewers attack it and a separate judge sort the valid points.
Copy this prompt
Here is my draft: [PASTE OR ATTACH]. Start 5 separate reviewers, each with one job: find factual errors, find weak arguments, find what I misunderstood, find what's missing, and check every number against a primary source with a link. Then, as a separate judge, list only the points that hold up, most important first, and say why you dropped the rest. Don't rewrite the draft yet.
Our own prompt, built around the lesson. We haven't run it ourselves. Text in [BRACKETS] is what you replace.
Where this comes from
Original post by @siliconvalleymm on X, Oct 22, 2026. We rewrote it in our own words. Check the original for the full context.
Read the originalGuides for this job
Turn a messy inbox into a 5-minute daily summary
Paste in today's emails and get back what needs a reply today, what can wait, what to ignore, and a one-line reply for each. It also flags emails that look like scams.