Value investing by committee: four AI agents, one verdict

What it does AI Berkshire is a set of 16 Claude Code skills—essentially structured prompt packs—that turn a single LLM session into a parallel research team. Each skill dispatches up to four independent agents, each adopting the lens of a different value-investing master (Buffett, Munger, Duan Yongping, or Li Lu), to research a company and deliver a forced pass/fail/gray-zone verdict with price targets and risk flags. It is not a standalone trading bot; it is configuration and methodology layered on top of Anthropic’s Claude Code.

The interesting bit The framework’s core trick is manufacturing disagreement. Instead of the usual LLM both-sidesism, the agents are instructed to contradict one another—Buffett might flag a “printing press” valuation while Li Lu rejects the same stock for uncertain management culture—so the user sees the tension rather than a smoothed-over summary. A “mirror test” rule even mandates rejection if the thesis can’t be explained in five sentences.

Key highlights

• 16 skills covering deep research, earnings review, industry screening, portfolio management, and “thinking tools” like Munger-style inversion.
• Four-agent parallel research with dedicated search and cross-validation per agent, coordinated by a Team Lead role.
• Rigid anti-bias guardrails: an 8-item veto checklist, anti-consensus checks, information-richness grades (A/B/C), and a “leave blank” principle that forces the model to admit uncertainty rather than hallucinate confidence.
• Financial rigor scripts using Python decimal.Decimal to cross-check market-cap and valuation figures against at least two independent sources.
• The author advertises a self-reported 2024 return of +69.29% and 2025 YTD return of +66.38% from a real brokerage account, with detailed outperformance tables versus the S&P 500 and Hang Seng.

Caveats

• This is fundamentally a collection of Markdown skill files and Python helper scripts for Claude Code; you bring your own subscription, API keys, and execution environment.
• The stellar track record is self-reported from a single unaudited account, so treat the advertised returns as anecdotal, not verified performance.
• The methodology is tightly coupled to a specific value-investing sect (Buffett/Munger/Duan/Li) and China/HK equity markets; growth investors or quant traders will find little here.

Verdict Worth a look if you already use Claude Code and want disciplined, reproducible equity research that forces conclusions instead of wall-of-text neutrality. Skip it if you’re looking for an autonomous trading system or a plug-and-play robo-advisor.