A dude from reddit conducted a cool experiment.
He took all 48 of Warren Buffett's letters to shareholders — that's almost 562 thousand words over 48 years. Loaded them into Claude Opus 4.6 and asked it to extract the main investing rules from them.
Then — the most interesting part. He gave the AI 50 companies, but removed all names. No hints — only financial indicators. And asked: "Choose the best ones according to Buffett's rules".
What happened:
→ 6 out of 10 of the AI's best picks are real companies from Buffett's portfolio
→ Berkshire Hathaway itself landed in 7th place — the AI didn't know it was Buffett's company
→ 13 out of 15 deliberately "bad" stocks were correctly filtered out
The only funny miss — Coinbase in 4th place. Buffett can't stand crypto, but Coinbase has a 39% margin, and by the numbers it looks excellent.
Architecture: Claude Code as orchestrator + 6 subagents. Applicable to any analysis where objectivity is needed.
discussion on reddit
