# ai-berkshire

> Value-investing research skills for Claude Code and Codex, based on four investors

- Repository: https://github.com/xbtlin/ai-berkshire
- Category: Agent skills & plugins
- Stars: 16,846 · Forks: 2,470 · Language: HTML · License: MIT
- Experience: Some experience
- Page: https://ir.adyog.com/repos/xbtlin/ai-berkshire/

AI Berkshire is a collection of skills, which are prompt-and-workflow files you install into Claude Code or Codex. Each one is a slash command such as /investment-research or /earnings-review. Together they cover company research, earnings reading, industry screening, portfolio review and tracking whether a buy thesis still holds.

The main idea is to avoid the balanced, noncommittal answer you get from asking a chatbot about a stock. The skills ask for a pass, fail or grey-zone verdict with price ranges. They split the analysis across four viewpoints that can disagree: business model, financial valuation, contrarian risk and long-term certainty. Team skills like /investment-team launch four agents in parallel. Each searches and scores on its own, and a team lead then combines the results.

The README also describes guardrails: a rating for how much information is available, a list of quick-reject red flags, a check against consensus thinking, and a Python helper that verifies figures such as market cap using exact decimals. The author reports personal trading returns and publishes a large index of generated reports. Those are the author's own claims and are not independently verified. The README warns that the deep skills use a lot of tokens.

## What it does

- **20 task-specific skills**: Skills are grouped into deep research, earnings analysis, industry screening, portfolio management and thinking tools. Examples include /quality-screen, /news-pulse, /thesis-tracker and /thesis-drift.
- **Four-perspective analysis**: Each company is judged through four investors' lenses that can conflict, then summarized with scores and a final verdict instead of a hedged essay.
- **Parallel multi-agent teams**: /investment-team and /earnings-team run four agents at once, each doing its own web research, with a team lead combining their conclusions.
- **Anti-bias and number checks**: Includes information-richness ratings, a contrarian failure-scenario check, an 8-item quick-reject list, and a script that checks market cap against price times shares.
- **Works with Claude Code and Codex**: Skills live as Claude Code command files, with scripts that generate and install matching Codex skills and optional Codex slash prompts. Install scripts exist for macOS, Linux and Windows.

## What you could build

- Run a full multi-agent research report on a listed company such as Tencent or Meituan.
- Screen a list of stocks quickly with the six-step buy checklist before spending tokens on deep research.
- Read a company's annual or quarterly report with the earnings-review skill, using only primary filings.
- Compare two versions of an investment thesis to see whether facts, valuation or just wording changed.
- Map an industry's supply chain and shortlist the companies worth a closer look.

## Who it's for

Suits individual investors comfortable with Claude Code or Codex who want a structured research workflow; you should already know basic investing terms, read the Chinese-language README, and treat the output as research input, not advice.

## Get started

Install Claude Code or Codex, clone the repo, and run the install script for your client. Then call a skill, for example /investment-research followed by a company name. Restart Codex after installing so it picks up the skills.

```
git clone https://github.com/xbtlin/ai-berkshire.git
```
```
cd ai-berkshire
./scripts/install-claude-commands.sh
```
```
/investment-research 腾讯
```

Tags: value-investing, claude-code, codex, stock-analysis, multi-agent