xbtlin/ai-berkshire

AI 时代的伯克希尔：基于 Claude Code / Codex 的价值投资研究框架。巴菲特·芒格·段永平·李录四大师方法论 + 多Agent并行研究。| AI-era Berkshire: a value investing research framework built for Claude Code / Codex. 4 masters' methodologies + multi-agent adversarial analysis.

What it solves

AI Berkshire addresses the lack of decision-grade quality in standard AI-generated investment analysis. While general LLMs often provide balanced but vague "on the one hand, on the other hand" responses, this framework enforces strict decision discipline, multi-perspective adversarial analysis, and financial precision to produce actionable investment research reports.

How it works

The project provides a collection of "Skills" (structured workflows) compatible with Claude Code and Codex. It transforms a single user into a virtual investment team through three layers:

  1. Skill Layer: 19 specialized entry points for tasks like deep research, earnings reviews, industry screening, and portfolio management.
  2. Agent Layer: For complex tasks, a "Team Lead" orchestrates four independent AI Agents representing the investment philosophies of Warren Buffett, Charlie Munger, Duan Yongping, and Li Lu. These agents independently search, analyze, and challenge each other's conclusions to eliminate blind spots.
  3. Tool Layer: A Python-based financial rigor tool (financial_rigor.py) that uses decimal.Decimal for precise calculations, cross-validates data from multiple sources, and performs market cap verification to prevent LLM hallucinations in financial math.

Who it’s for

It is designed for value investors and financial analysts who want to systematize the methodologies of legendary investors using AI to increase research depth and efficiency without sacrificing rigor.

Highlights

  • Adversarial Analysis: Uses four distinct master-investor perspectives to create tension and conflict in analysis rather than a single consensus.
  • Anti-Bias Mechanisms: Includes information richness ratings (A/B/C), Munger-style inversion (forcing the AI to imagine failure scenarios), and a "Mirror Test" for final decision discipline.
  • Financial Rigor: Implements a dedicated Python tool to verify market caps and valuations, ensuring calculations are mathematically exact rather than estimated by the LLM.
  • Multi-Agent Parallelism: The /investment-team skill runs four agents in parallel, quadrupling the search volume and information sources compared to a single prompt.
  • Comprehensive Skill Set: Covers the entire investment lifecycle from initial industry funneling and quality screening to deep-dive research and post-purchase thesis tracking.