AI Berkshire (ai-berkshire) is an MIT-licensed collection of value-investing research skills for Claude Code and Codex that turns the methodologies of Warren Buffett, Charlie Munger, Duan Yongping and Li Lu into reproducible, multi-agent fundamental analysis for individual investors and small research teams.

What it is

AI Berkshire is a skill collection that runs inside Claude Code and Codex. It organises 20 skill entry points spanning deep research, financial report analysis, industry screening, position management and thinking tools, and layers a set of team-type skills on top of them, including /investment-team. Supporting code lives in files such as tools/financial_rigor.py, and the project ships documentation in Chinese, English (README_EN.md) and Japanese (README_JA.md). It sits in the AI & Machine Learning / AI Development Platforms category, is written primarily in HTML, carries 16,427 stars, 2,451 forks and 38 open issues, and was last pushed on 11 September 2026.

The specific thing it replaces is the unstructured answer a general model gives when asked whether a stock is worth buying. That answer balances one side against the other and closes with a warning to judge for yourself, which leaves nothing to act on. AI Berkshire forces a verdict — 通过 (pass), 不通过 (fail) or 灰色地带 (grey zone) — attaches specific price bands and tiered suggestions to it, and applies a mirror test: a thesis that cannot be stated in five sentences is a pass on buying, with no exceptions. The README states the difference plainly: one person plus Claude Code or Codex equals a research team.

Key capabilities

• Forced conclusions replace balanced commentary. A Pinduoduo analysis outputs an aggressive entry of 20% at $95-105, a steadier entry at $85-95 pending clarity on buyback policy, and a conservative stance of no position because the ten-year certainty standard is unmet.
• Four masters score each company independently and are allowed to contradict each other. For Pinduoduo: Duan Yongping on business model at 3.7/5, Buffett on valuation at 4.4/5 with ex-cash PE of 6.3x, Munger on inversion at 3.5/5, and Li Lu on long-term certainty at 2.0/5.
• /investment-team launches four independent agents on one company at the same time. Each searches the web, cross-verifies data and reaches its own conclusion; a Team Lead then synthesises the four results.
• Anti-bias controls are built into the workflow: A/B/C information-richness grading (Pop Mart is rated B, with inferred metrics annotated by confidence), Munger-style inversion that lists five failure scenarios with probabilities, an eight-item instant-veto checklist, a contrarian check, and a deliberate blanks principle that marks grey zones rather than disguising guesses as certainty.
• Numeric rigour is enforced through tooling rather than mental arithmetic. python3 tools/financial_rigor.py verify-market-cap --price 510 --shares 9.11e9 --reported 4.65e12 --currency HKD reconciles a Tencent market-cap figure to a 0.08% deviation, all calculations use Python decimal.Decimal rather than float, and key figures need at least two independent sources.
• Outputs are reproducible. Seven companies were screened against one shared checklist covering circle of competence, business quality, moat, management and margin of safety, producing 4.7 for 茅台 and 腾讯, conditional passes for 英伟达 (4.3), 美团 (4.0) and 快手 (4.0), and grey-zone results for 拼多多 (3.8) and 泡泡玛特 (3.7).
• A published corpus accompanies the framework. reports/README.md indexes 2,346 reports across 110 companies and 23 topics, updated to 12 September 2026.

Who uses it and how

• A single investor working alone, using the skills as a substitute for an analyst team and running /investment-team company by company.
• Small teams that need aligned research: identical inputs produce consistent structure and depth, so seven companies can be compared on the same scoring standard, or one company can be re-analysed six months later and diffed against the earlier report.
• Investors focused on Chinese and Asian equities, the area the topics list emphasises through china-stock, berkshire-hathaway and charlie-munger, with worked examples on 茅台, 腾讯, 美团, 快手, 拼多多 and 泡泡玛特 alongside 英伟达.
• Readers who consume the output rather than run the framework, following the indexed report corpus or the curated selection published on the WeChat account 复利炼丹炉.
• Users evaluating the stated track record, which reports 2024 returns of +69.29% and 2025 returns of +66.38% against the Hang Seng, S&P 500, CSI 300 and Nasdaq, with screenshots from a Futu account and a cumulative two-year figure above ¥1.46 million.

Getting started

Setup is documented in the README's Quick Start section, which covers installing the skill collection into Claude Code or Codex. The registry entry names no package or container image, so the repository README is the distribution channel.

How it compares

The facts provided name no paid products that AI Berkshire replaces, and no directly comparable research framework either, so within this registry it stands alone. Claude Code and Codex appear here as the host agent runtimes it depends on rather than as rivals it displaces.

When to use it — and when not

Adopting it means operating an agent runtime: a Claude Code or Codex environment, a Python 3 interpreter for tools/financial_rigor.py, and network access for the four parallel agents to search and verify. Anyone who wants a hosted interface, a managed service or a single-click product should look elsewhere, and the performance figures are self-reported from a personal brokerage account with the README itself noting that past returns do not indicate future results. The repository is also Chinese-first, with English and Japanese translations alongside, and it is classified under HTML despite being an agent skill collection, so contributors should expect to read the Chinese documentation for the fullest detail.

# 市值手算校验：股价 × 总股本，与报告数据对比
python3 tools/financial_rigor.py verify-market-cap \
  --price 510 --shares 9.11e9 --reported 4.65e12 --currency HKD
# ✅ 验证通过, 偏差仅 0.08%

npm install -g @anthropic-ai/claude-code

# macOS / Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh

# 或使用 npm
npm install -g @openai/codex

# 或使用 Homebrew
brew install --cask codex

# 验证安装
codex --version

claude --dangerously-skip-permissions

##