AI Berkshire: AI Investment Research Skills
/ai-berkshire
AI Berkshire is a skill collection for Claude Code and Codex that turns value-investing frameworks from Warren Buffett, Charlie Munger, Duan Yongping, and Li Lu into structured, multi-agent research workflows. The README backs it with two years of live returns but calls the output research, not investment advice. Reach for it if you read primary filings and want a sourced conclusion instead of hedged prose; skip it if you just want stock picks.
What is AI Berkshire?
AI Berkshire is a collection of about 20 investment-research skills that install into Claude Code or Codex and structure how the AI investigates a company. Each skill runs a fixed pipeline — data, business quality, moat, management, valuation — instead of one freeform prompt, and the flagship /investment-team skill runs four agents arguing from Buffett's, Munger's, Duan Yongping's, and Li Lu's views before a team-lead agent synthesizes a verdict.
Why Direct AI Prompts Fall Short for Investing
Ask Claude directly whether a stock is worth buying and the README says you get a hedged "on one hand, on the other hand" answer that ends in "investing carries risk, decide for yourself" — accurate-sounding, but nothing to act on. AI Berkshire's stated fix isn't better prompting, it's forcing structure: every skill outputs a pass/fail/gray-zone call with a specific price range instead of a summary, and a single context window doesn't get to pretend it covered a topic from every angle. The README also flags LLM mental math as a real failure mode — a misplaced decimal in a P/E ratio or a mixed-up HKD/RMB market-cap unit can flip a conclusion — which is why the framework routes calculations through a Python script instead of having the model estimate them.
Core Capabilities for Structured Investment Research
- ✓Four-master adversarial analysis: /investment-team runs Buffett, Munger, Duan Yongping, and Li Lu as four independent agents that search and score separately before a team-lead agent reconciles their disagreement — the README's own PDD example has the four scoring 4.4, 3.7, 3.5, and 2.0 on the same company.
- ✓Structured anti-bias mechanisms: an A/B/C information-richness rating flags thin data, a Munger-style inversion prompt asks what would kill the thesis, an 8-item quick-veto checklist can reject a company on management integrity alone regardless of valuation, and the framework is told to output 'gray zone' rather than fake certainty.
- ✓Financial-rigor tooling: tools/financial_rigor.py runs market-cap and valuation checks (verify-market-cap, verify-valuation), cross-validates a metric across multiple sources, and runs a Benford's-law check on financial figures — all in Python's decimal.Decimal rather than float.
- ✓Reproducible checklist output: /investment-checklist and /quality-screen apply the same rubric across multiple tickers side by side, which is how the README's own comparison table scores seven companies (Moutai, Tencent, Nvidia, Meituan, Kuaishou, PDD, Pop Mart) on one 5-star scale.
- ✓Three-layer design: a skill layer of roughly 20 entry points for specific research jobs, an agent layer where team skills coordinate the four master-perspective agents, and a tool layer that handles precise calculation and cross-checking underneath both.
- ✓Dual-client packaging: the same canonical workflow ships as Claude Code commands (skills/*.md) and as Codex skill packages (codex-skills/*/SKILL.md) generated from those same source files via scripts/sync-codex-skills.py.
Proven Track Record and Performance
- ✓The README backs the framework with two years of the maintainer's own live account: +69.29% in 2024 and +66.38% in 2025, screenshotted from a Futu Securities account, versus the S&P 500's +23.31% and +16.39% and the Hang Seng's +17.67% and +27.77% over the same two years.
- ✓Financial-rigor tooling catches a real class of AI error: the README's own worked example is a Tencent market-cap mix-up between HKD and RMB units, caught by running the number through verify-market-cap rather than trusting the model's mental math.
- ✓MIT license — no restriction on forking the skills, editing the master prompts, or wiring in your own data sources.
- ✓Cross-client support: the same skills install into both Claude Code and Codex, generated from one source (skills/*.md) rather than maintained twice.
Current Limitations and Future Vision
- △The headline returns are a single account's two-year track record, not a backtest across market cycles — the README's own disclaimer says historical returns don't represent future performance.
- △Deep-research skills run multiple rounds of search, cross-validation, and multi-agent synthesis, which the README says costs meaningfully more tokens than a single prompt, and the maintainer's stated view is not to downgrade the model just to save cost on decisions that matter.
- △The roadmap section lists historical backtesting of AI reports against actual price performance, a macro-cycle framework, and MCP-based real-time data (Wind/Bloomberg/Yahoo Finance) as unbuilt future work, not shipped features.
- △No version number or release history appears in the facts available here, so there's no documented way to tell how stable any given skill's output format is release to release.
- △The project explicitly frames itself as a research aid, not a decision engine — the README's disclaimer states it's for learning and research only and isn't investment advice, so treat any pass/fail/gray-zone call as an input, not an answer.
Setting Up AI Berkshire
Install an AI client first: `npm install -g @anthropic-ai/claude-code` for Claude Code, or Codex via `npm install -g @openai/codex` (also available as a curl script or `brew install --cask codex`, verified with `codex --version`). Then clone the framework itself — `git clone https://github.com/xbtlin/ai-berkshire.git` — and run the matching installer: `./scripts/install-claude-commands.sh` copies the skills into Claude Code's global commands directory, while `./scripts/install-codex-skills.sh` installs them to `~/.codex/skills` (an optional `./scripts/install-codex-prompts.sh` adds Codex slash-prompt compatibility). Windows users run the equivalent `.bat` scripts from PowerShell or Command Prompt. Because these skills call tools frequently, Claude Code prompts for permission on every call by default — the README notes you can start with `claude --dangerously-skip-permissions` to skip that, but only in an environment and repo you actually trust, since it turns off tool-approval protection entirely.
Exploring the 20+ Investment Skills
Skills are invoked as slash commands once installed: `/investment-research 腾讯` runs the full seven-module single-company workup, `/investment-team 美团` launches the four-agent parallel version, and `/investment-checklist 腾讯, 阿里巴巴, 美团` runs the six-gate buy checklist across several tickers at once. Earnings get their own pair — `/earnings-review` for a primary-source read of a single filing, `/earnings-team` for a four-master read that ends in a publishable WeChat-style article. Industry-level research runs through `/industry-research` for a supply-chain map, or `/industry-funnel`, which the README describes as narrowing a market from roughly 30-60 candidates down to a final 3. Position management covers `/portfolio-review`, `/thesis-tracker`, `/thesis-drift` for comparing two theses over time, and `/news-pulse` for a fast attribution pass when a stock moves sharply. In Codex, the same skills are triggered by describing the task in natural language after a restart, or via `/prompts:investment-research` if the optional slash-prompt layer is installed. The README's own cost guidance: run the cheap screens (`/quality-screen`, `/news-pulse`) first, and only reach for the full multi-agent skills once a candidate looks worth the token spend.
Other AI Tools for Financial Analysis
Frequently Asked Questions
AI Berkshire is built to run inside Claude Code and Codex specifically, installing as Claude Code commands or Codex skills. The README doesn't list a compatibility matrix of individual model versions beyond those two clients — check each client's own supported-model docs for specifics.
AI Berkshire runs financial calculations through tools/financial_rigor.py using Python's decimal.Decimal instead of letting the model estimate, cross-validates key figures across at least two independent sources, and includes a Benford's-law check for anomalies in financial figures, per the README.
AI Berkshire structures its analysis around four value-investing perspectives named in the README: Warren Buffett (moat and valuation), Charlie Munger (inversion and bias-checking), Duan Yongping (business-model quality), and Li Lu (long-term civilizational trends).
AI Berkshire's own README disclaimer states the project is for learning and research purposes only, does not constitute investment advice, and tells users to do their own due diligence — so it isn't positioned as a direct decision-maker.
The README doesn't give a dollar or token figure — it says deep-research skills run multiple rounds of search, cross-validation, and multi-agent synthesis, which costs meaningfully more tokens than a single prompt, and recommends running cheaper screens like /quality-screen or /news-pulse first.
AI Berkshire is released under the MIT License, as listed on its GitHub repository.
Best use cases
- •Building a documented research memo before buying or adding to a position, since /investment-research and /investment-team both end in a structured pass/fail/gray-zone call with a price range instead of a summary paragraph.
- •Screening a sector or index down to a shortlist — /industry-funnel narrows a market from a broad candidate set to a handful of finalists, and /quality-screen applies 7 hard filters across individual stocks, an industry, an index, or a theme.
- •Reading an earnings report against primary filings instead of secondhand summaries with /earnings-review, or turning that read into a publishable write-up with /earnings-team.
- •Tracking whether an existing investment thesis still holds — /thesis-tracker follows a position over time and /thesis-drift diffs two versions of a thesis to separate a genuine fact change from a valuation change or just different wording.
- •Getting a fast read on why a stock moved sharply, with /news-pulse's 10-15 minute attribution pass, before deciding whether the move actually warrants a full deep-research run.
Who should try it — and who should skip
Try AI Berkshire if you already do fundamental research by hand — reading primary filings, building your own valuation models — and want an AI workflow that forces a specific, sourced conclusion instead of hedged prose, especially if you're comfortable running Claude Code or Codex from a terminal and paying for the token cost of multi-agent analysis. Skip it if you want investment recommendations rather than a research process: the README is explicit that this is a learning and research tool, not an advisor, and every output still needs your own judgment and due diligence on top of it.
Related repositories
Still deciding about ai-berkshire?
One click hands the question to an AI along with this page — see what it says about ai-berkshire.