After making +1.46 million in two years, someone open-sourced their investment research system. The project is called AI Berkshire, and the real trading results are also publicly available:
2024: +69.29%
2025: +66.38%

The core isn’t “AI stock picking” — it’s about breaking down the investment frameworks of Buffett, Munger, Duan Yongping, and Li Lu into executable, reviewable research workflows.

It conducts company research from perspectives like business model, moat, cash flow, valuation, margin of safety, management, and risk factors, then uses Claude Code / Codex to have multiple agents work in parallel on financial analysis, valuation judgment, cross-validation, and investment memos.

The most valuable takeaway from this project is that it doesn’t treat AI as an “answer machine,” but instead organizes the judgment paths in real investment research into a repeatable system.

For individual investors, this is likely more valuable than simply asking AI, “Can I buy this stock?”

Project: https://github.com/xbtlin/ai-berkshire
Not investment advice, purely for sharing and learning.

xbtlin/ai-berkshire

Source: https://github.com/xbtlin/ai-berkshire
中文 | English | 日本語

GitHub Trending (https://trendshift.io/repositories/63696)

AI Berkshire - A Value Investing Research Framework for the AI Era

“Price is what you pay, value is what you get.” — Warren Buffett

Redefining the depth and efficiency of investment research with AI.

AI Berkshire is a collection of investment research Skills that works with both Claude Code and Codex. It systematically structures the methodologies of four value investing masters—Buffett, Munger, Duan Yongping, and Li Lu—into professional-grade investment research driven by AI agents.

One person + Claude Code / Codex = an entire research team.

Real Track Record · Why You Can’t Just Ask AI Directly · Skills Overview (19) · Quick Start · Real Research Reports · Design Philosophy · WeChat Account

Real Track Record

This isn’t just theoretical. The framework is backed by a real-money, verified investment system.

2024 Full Year Return: +69.29%

2025 Full Year Return: +66.38%

Comparison with Major Indices

2024 Excess Return: Outperformed S&P 500 by 46 percentage points, outperformed Hang Seng Index by 52 percentage points
2025 Excess Return: Outperformed S&P 500 by 50 percentage points, outperformed Hang Seng Index by 39 percentage points

Cumulative real account return over two years exceeds 1.46 million yuan, significantly outperforming major global indices for two consecutive years.

Disclaimer: Past performance does not guarantee future results. Screenshots are from a real Futu Securities account.

Select Research First Published on WeChat

This repository contains the complete framework and all reports. The WeChat account features selected content—in-depth research on stocks truly worth betting on, along with my own judgments and trade-offs beyond the reports:

复利炼丹炉 (Compound Interest Furnace) — Using AI to refine the pill of research.

Why You Can’t Just Ask AI

You can certainly ask Claude directly: “Help me analyze whether Pinduoduo is worth buying.” You’ll get a balanced analysis that goes “on one hand… on the other hand…”, ending with “Investing involves risks, please judge for yourself.”

That kind of analysis looks right but can’t actually be used to make decisions.

AI Berkshire solves not whether to analyze, but analysis quality and decision discipline. Here are the core differences:

1. Forces a Conclusion, No Hedging

When you ask AI directly, you get a wishy-washy analysis that tries to please everyone. AI Berkshire forces an output: Pass/Fail/Gray Zone, with specific price ranges and tiered recommendations.

Generic AI answer: “Pinduoduo has growth potential but also faces competitive pressure, investors need to weigh…”

AI Berkshire output:

Strategy	Recommendation	Price Range
Aggressive	Can build 20% position at current price	$95–105
Conservative	Build position after buyback policy is clear	$85–95
Defensive	Doesn’t meet 10-year certainty standard, wait	—

Mirror Test: Can’t explain it in 5 sentences = don’t buy. No exceptions.

2. Four Masters’ Perspectives in Conflict, Not a Single Analysis

It’s not just “analyze with Buffett’s method.” Four viewpoints produce real tension and friction—using Pinduoduo as an example:

• Duan Yongping (Business Model): Good business, C2M model hard to replicate → Score 3.7/5
• Buffett (Financial Valuation): P/E ex-cash only 6.3x, cash machine → Score 4.4/5
• Munger (Inverse Thinking): Moat is shallower than expected; Douyin hit 4 trillion GMV in 3 years → Score 3.5/5
• Li Lu (Long-term Certainty): Management culture has issues, uncertain after 10 years → Score 2.0/5

Buffett says “It’s really cheap,” Li Lu says “If uncertain, don’t buy” — this conflict is the real state of investment decisions. A single prompt can’t create this multi-perspective conflict, which is precisely the key to avoiding blind spots.

3. Structured Anti-Bias Mechanisms

AI’s most dangerous output isn’t a wrong answer, but one that looks right but falls apart under scrutiny. AI Berkshire embeds multiple layers of “scam protection” in its workflow:

4. Financial Data Precision

LLM mental math is unreliable. A decimal point error in P/E or mixing up HK dollars with RMB for market cap can lead to wrong investment decisions.

Real Case: When analyzing Tencent, market cap data from different sources came in “HKD billions” and “RMB billions” units. AI Berkshire’s handling:

# Market cap manual verification: stock price × total shares, compare with reported data
python3 tools/financial_rigor.py verify-market-cap \
  --price 510 --shares 9.11e9 --reported 4.65e12 --currency HKD
# ✅ Verification passed, deviation only 0.08%

All calculations use Python decimal.Decimal (exact decimal), not float. Critical data is cross-verified from at least two independent sources.

5. Reproducible Research Process

Ask AI directly, and every time you get a different format, depth, and scope—today’s analysis of Tencent has a moat score, tomorrow’s analysis of Meituan might forget it.

AI Berkshire ensures: Same input → consistent structure and depth of output. This means you can:

• Compare 7 companies horizontally with identical scoring criteria
• Re-analyze the same company six months later and directly compare changes
• Align research results across team members

Real output—7 companies screened using the same Checklist criteria:

Company	Pass?	Circle of Competence	Good Business	Moat	Management	Margin of Safety	Overall
Moutai	✅ Pass	★★★★★	★★★★★	★★★★★	★★★☆☆	★★★★☆	4.7
Tencent	✅ Pass	★★★★☆	★★★★★	★★★★★	★★★★★	★★★★☆	4.7
NVIDIA	✅ Conditional	★★★★☆	★★★★★	★★★★★	★★★★★	★★★☆☆	4.3
Meituan	✅ Conditional	★★★★☆	★★★★☆	★★★★☆	★★★★☆	★★★★☆	4.0
Kuaishou	✅ Conditional	★★★☆☆	★★★★☆	★★★★☆	★★★★☆	★★★★★	4.0
Pinduoduo	❓ Gray	★★★★☆	★★★★☆	★★★☆☆	★★★☆☆	★★★★★	3.8
Pop Mart	❓ Gray	★★★☆☆	★★★★☆	★★★★☆	★★★★★	★★★☆☆	3.7

6. Multi-Agent Parallelism = Multiplied Research Depth

/investment-team launches 4 independent agents to research a company simultaneously. Each agent independently searches the web, cross-validates data, and gives its own conclusion. This isn’t splitting one prompt into four paragraphs—it’s 4 “analysts” each doing a complete study, then a Team Lead synthesizes.

When a person asks AI directly, the context window is one. With 4 agents in parallel, you get 4x the search volume, 4x the information sources, and 4 independent perspectives.

One-Sentence Summary

An ordinary person asking AI gets “an analysis that looks right”; using AI Berkshire, you get “an investment research report you can actually make decisions on.”

Overall Architecture

Three-layer design philosophy:

• Skill Layer: Abstracts “what you want to do” into 19 clear entry points—deep research, financial analysis, industry screening, portfolio management, thinking tools. Choose by scenario.
• Agent Layer: Team skills (e.g. /investment-team, /earnings-team) use a Team Lead to schedule 4 master-perspective agents in parallel—each independently searches, judges, and challenges the others, then synthesizes a conclusion. Lightweight skills skip this layer and go directly to tools for quick in-and-out.
• Tool Layer: Precise calculations, real-time retrieval, report spot-checking—ensuring the data rigor of every report is verifiable.

Skills Overview (19)

🔬 Deep Research

📊 Financial Analysis

🏭 Industry Screening

📈 Portfolio Management

🧠 Thinking Tools

Quick Start

Cost & Model Selection

Deep research Skills perform multiple rounds of research, cross-validation, and multi-agent synthesis by default, so token consumption is higher. This is the trade-off for more complete business, financial, industry, and risk analysis.

For high-risk, high-importance judgments in real investment decisions, the maintainer’s view is that stronger models are generally more likely to provide better analysis ROI; it’s not recommended to sacrifice critical judgment quality just to save on model costs. Lighter models are better suited for initial screening, summarization, or low-risk questions. For analyses involving moat, valuation, management, and risk cross-judgment, expect analysis quality to be more dependent on model capability.

To control costs, adjust the workflow first instead of expecting full deep research to become cheap: use /quality-screen to quickly eliminate companies, or use /news-pulse for price move attribution. Only run /investment-research or /investment-team when results justify deeper work.

1. Install AI Client

This repository maintains the same canonical workflow and provides commands for both Claude Code and Codex skills. Install according to the client you use.

Claude Code users:

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

Codex users:

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

# or using npm
npm install -g @openai/codex

# or using Homebrew
brew install --cask codex

# Verify installation
codex --version

Windows users can use the official PowerShell installation command:
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex".

If codex --version outputs a version number, you can proceed to install this project’s Codex skills.

Reducing Authorization Prompts

These skills frequently call tools. Claude Code requests authorization for each one by default. This behavior comes from Claude Code’s client permission mechanism and is not something this repository can change by default.

If you trust the current workflow and are operating in a trusted environment, you can launch Claude Code with the skip-permissions mode:

claude --dangerously-skip-permissions

Note: This mode disables Claude Code’s tool approval protection. Only use it if you trust the repository, commands, and working directory.

2. Install Skills

Claude Code users (macOS / Linux):

# Clone the repository
git clone https://github.com/xbtlin/ai-berkshire.git

# Copy skills to Claude Code's global commands directory
cd ai-berkshire
./scripts/install-claude-commands.sh

Claude Code users (Windows PowerShell / Command Prompt):

git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire
.\scripts\install-claude-commands.bat

Codex users (macOS / Linux):

# Clone the repository
git clone https://github.com/xbtlin/ai-berkshire.git

# Generate and install Codex skills to ~/.codex/skills
cd ai-berkshire
./scripts/install-codex-skills.sh

# Optional: Install Codex slash prompts to ~/.codex/prompts
# For a closer /investment-research experience to Claude Code
./scripts/install-codex-prompts.sh

Codex users (Windows PowerShell / Command Prompt):

git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire
.\scripts\install-codex-skills.bat

REM Optional: Install Codex slash prompts
.\scripts\install-codex-prompts.bat

The repository maintains three sets of entry points: skills/*.md are the Claude Code command source files; codex-skills/*/SKILL.md are Codex skill packages, generated by scripts/sync-codex-skills.py from skills/*.md; codex-prompts/*.md are an optional Codex slash prompt compatibility layer.

3. Usage

In Claude Code, call directly:

# Deep research
/investment-research Tencent
/investment-team Meituan
/management-deep-dive Wang Xing Meituan
/private-company-research SpaceX
/deep-company-series Pinduoduo

# Financial analysis
/earnings-review Tencent 2025Q4
/earnings-team PDD 2025 Annual Report

# Industry screening
/industry-research Nuclear Power
/industry-funnel AI Computing
/quality-screen Hang Seng Index Constituents
/bottleneck-hunter AI Infrastructure
/investment-checklist Moutai, NVIDIA, Apple

# Portfolio management
/portfolio-review Tencent 30%, Meituan 20%, Moutai 20%, Cash 30%
/thesis-tracker Pinduoduo
/thesis-drift Pinduoduo reports/Pinduoduo-thesis-2025Q4.md reports/Pinduoduo-thesis-2026Q1.md
/news-pulse Tencent

# Thinking tools
/dyp-ask Where exactly is Pinduoduo's moat?
/wechat-article Meituan

In Codex, after installation, restart Codex and describe the task by skill name, for example:

Use investment-research to research Tencent
Use earnings-review to analyze PDD 2025 annual report
Use industry-funnel to screen AI Computing
Use bottleneck-hunter to scan AI infrastructure bottlenecks
Use thesis-drift to compare two Pinduoduo investment theses
Use wechat-article to write a Meituan research article

If you installed Codex slash prompts, after restarting Codex you can also search for these prompts in the / menu. Codex’s official custom prompt entry usually appears as prompts:, for example:

/prompts:investment-research Tencent

Detailed Skill Descriptions

1. /investment-research — Four Masters Comprehensive Analysis

The most comprehensive deep research framework for a single company. Executes in seven modules sequentially:

Data Collection → Business Essence (Duan Yongping) → Moat (Buffett) → Inverse Thinking (Munger) → Management Assessment (Duan Yongping + Buffett) → Civilization Trend (Li Lu) → Valuation & Margin of Safety

Key Features:

• AI research bias awareness mechanism (A/B/C information richness rating)
• Multi-source cross-validation of key data (manual market cap verification, at least 2 independent sources)
• “Follow-up questions” from the four masters throughout the text
• Three-scenario valuation (optimistic/neutral/pessimistic) + Reverse DCF

Output Example Excerpt:

Summary Decision Memo

Dimension	Conclusion	Confidence
Business Quality (Duan Yongping)	Excellent: Platform business, bilateral network effects, marginal cost approaches zero	★★★★★
Moat (Buffett)	Wide and widening: Triple overlay of network effects + switching costs + economies of scale	★★★★☆
Management (Duan Yongping + Buffett)	Good: Founder at helm, strong capital allocation discipline	★★★★☆
Biggest Risk (Munger)	Regulatory policy uncertainty, new business losses dragging down overall profit	★★★☆☆
Civilization Trend (Li Lu)	Aligns with digital consumption trends, but not a “civilization-level paradigm shift”	★★★★☆
Valuation (Buffett + Duan Yongping)	Current P/E 18x, at historical median on the low side, some margin of safety	★★★★☆

Duan Yongping: “The essence of this business is connecting consumers and merchants, earning money from efficiency gains. The sign of a good business: more users → more merchants; more merchants → more users. Once the flywheel spins, it’s hard to stop.”

Munger: “Think inversely—if this company disappeared tomorrow, what would users and merchants do? If the answer is ‘find a replacement quickly,’ the moat isn’t deep enough. If the answer is ‘life would become very inconvenient,’ that’s worth attention.”

2. /investment-team — Multi-Agent Research Team

Launches 4 AI agents to research in parallel, simulating a real research team collaboration. Each agent independently searches, analyzes, and scores, then a Team Lead synthesizes the final judgment.

Output Example Excerpt:

One-Sentence Conclusion

Meituan is the absolute leader in China’s local life services, with a moat built on multiple network effects. Current valuation is at historically low levels, with significant long-term investment value. Recommend accumulating positions on dips.

Four-Dimension Score Summary

Dimension	Framework	Score	Core Judgment
Business Model & Moat	Duan Yongping	★★★★☆	Strong bilateral network effects, delivery + in-store form a flywheel
Financials & Valuation	Buffett	★★★★☆	Core business margins continually improving, valuation at historical lows
Industry & Competition	Munger	★★★☆☆	Douyin encroaching on in-store business, competitive landscape may deteriorate
Risk & Management	Li Lu	★★★★☆	Wang Xing has excellent strategic vision, but new business cash burn is a concern

Overall Score: 3.8 / 5

Investment Recommendation

Strategy	Recommendation	Price Range (HKD)
Aggressive	Can build 30% position at current price	120–140
Conservative	Build position after pullback to 100–110	100–120
Defensive	Wait for quarterly earnings to confirm margin trend before entering	<100

3. /investment-checklist — Buffett’s Pre-Purchase Checklist

Six-level quick screen to decide in 10 minutes whether a company is worth deeper research:

Level 1: Circle of Competence (Can I understand it?)
    ↓ Pass
Level 2: Good Business (What are the economic characteristics?)
    ↓ Pass
Level 3: Moat (How deep is the competitive advantage?)
    ↓ Pass
Level 4: Management (Trustworthy?)
    ↓ Pass
Level 5: Margin of Safety (Is the price cheap?)
    ↓ Pass
Level 6: Decision Discipline (Is it rational or FOMO?)
    ↓ Pass
✅ Mirror Test

Supports multi-company comparison—screen multiple candidates at once:

/investment-checklist Tencent, Alibaba, Meituan, Pinduoduo

Output Example Excerpt:

Mirror Test

“I’m buying Tencent at 380 HKD because:

1. The essence of this business is social network + digital content platform, I understand it.
2. Its moat is 1.2 billion users’ social relationship chain, and it’s widening.
3. Management Pony Ma is low-key and pragmatic, capital allocation is excellent, trustworthy.
4. Current price is approximately 80% of intrinsic value, some margin of safety.
5. Even if I’m wrong, downside risk is manageable because net cash on books exceeds 200 billion, game cash flow is strong.“

✅ Passes Mirror Test

Can’t explain in 5 sentences = don’t buy. No exceptions.

4. /industry-research — Full Industry Chain Panoramic Scan

Starting from an investment theme, complete a panoramic study of the industry chain:

Build Investment Logic Chain → Industry Chain Panorama → Global Listed Company Scan → Four-Master Analysis of Key Players in Each Link → Portfolio Allocation Recommendation

Output Example Excerpt:

Investment Logic Chain: Nuclear Power

Underlying Trend: AI data center power demand surge + Carbon neutrality goals
→ Leads to: Soaring demand for stable, clean baseload power
→ Creates: Certain demand for nuclear restart/new builds/SMR
→ Benefits: Uranium mining → Fuel processing → Equipment manufacturing → Operators

Recommended Portfolio

Layer	Allocation	Ticker	Link	Core Logic
Core	50%	CGN Power, Cameco	Operation + Uranium	Highest certainty
Satellite	30%	China Nuclear Power, Dongfang Electric	Operation + Equipment	Beneficiary of import substitution
Option	15%	NuScale, Nano Nuclear	SMR	High risk, high optionality
ETF	Alternative	URA, URNM	Full chain	Lazy option

5. /industry-funnel — Industry Funnel Screening

Starting from an industry/direction, full market → ≤ 10 → 3 layer-by-layer selection:

Full market scan (union of activity + gainers + top 30 by market cap, 30–60 companies)
    ↓
5 hard value investing criteria → rough screen ≤ 10 companies
    ↓
Detailed analysis (300–500 words each) → fine screen ≤ 10 companies
    ↓
Final selection (by portfolio complementarity, not top 3 by score)
Four-master deep analysis of 3 companies (800–1200 words each)
    ↓
Recommended portfolio (Core / Satellite / Option) + Operation signals

Key Features:

• Each layer has clear keep/discard criteria; eliminated stocks come with reasons (no black box)
• Final 3 selected by “portfolio complementarity” (high certainty + medium optionality + high optionality), not ranked top 3 by score
• Mandatory “future IPO candidates” column avoids missing key private market players
• AI bias awareness mechanism: addresses leader bias / English bias / story bias / listing bias

Difference from /industry-research:

• industry-research focuses on industry chain structure and panorama (sliced by link)
• industry-funnel focuses on stock selection funnel (from full market down to 3)

Real Test: AI Industry 4 Sub-tracks in Parallel (2026-05-09):

Key Discovery: The biggest winners in AI applications are not AI-native companies, but established giants with distribution + data + workflow integration. This echoes the historical pattern of the 1995–2000 internet bubble (Amazon and Apple won; Pets.com lost).

Full reports: AI Compute · AI Models · AI Applications · AI Infrastructure Power

6. /private-company-research — Unlisted Company Deep Research

A “detective-style” research framework designed for information-scarce unlisted companies:

Key Differentiators:

• Financial data piecing together: Combine from prospectus, parent company financials, fundraising news, industry data from multiple sources
• Confidence labeling: Each data point labeled 🟢 High / 🟡 Medium / 🔴 Low confidence
• Multi-method valuation cross-check: Fundraising valuation + comparable company + DCF + endgame back-calculation
• Exit path analysis: Full assessment of IPO/M&A/secondary transfer

Output Example Excerpt:

Company Snapshot: SpaceX

Item	Content
Latest Valuation	~$350B (2025 secondary market) 🟡
Estimated Revenue	~$13B (2024) 🟡
Starlink Users	4M+ (end of 2024) 🟢
Launches	100+ / year (2024) 🟢

Valuation Judgment

Method	Valuation Range	Notes
Latest Fundraising	$350B	Secondary market quote, includes liquidity premium
Comparable Company	$200–280B	Benchmarking telecom + aerospace + defense
DCF (Neutral)	$250–350B	Assumes Starlink $30B revenue by 2027
Endgame Backward	$400–600B	Assumes Starlink becomes global telecom infrastructure

Composite reasonable valuation range: $250B – $400B

7. /news-pulse — Price Move Attribution

An intelligence response skill designed for “quickly figuring out what happened when a stock jumps or drops.” Not deep research—it’s a 10-15 minute rapid attribution to avoid panic from rumors or blindly stopping losses when positions move.

Key Differentiators:

• 4-dimensional parallel reconnaissance: Company events / Regulatory policy / Industry rivals / Market sentiment (sell-side + influencers + southbound flows)
• Attribution before listing: Not just listing all news, but judging “which event justifies this price move”
• Mandatory nature judgment: Value event / Sentiment fluctuation / True cause unknown / Mixed—“true cause unknown” is the most valuable output (possible insider front-running)
• Clear action suggestions: Trigger deep research? Need to review thesis? Observe only?

Difference from Other Skills:

Output Example Excerpt (Tencent 4/17–5/01 real test, 14 days -10.47%):

One-Sentence Attribution

About 70–80% of this -10.47% decline was driven by liquidity+ sentiment factors (buyback quiet period + southbound reduction + sector beta + AI narrative capture), and 20–30% by digestion of the doubled AI investment. No negative fundamentals, sell-side maintains consensus buy. Nature is a “liquidity + sentiment-driven correction,” not a value event.

Attribution Table

Candidate Explanation	Estimated Contribution	Confidence
Buyback quiet period disappearing (structural, before 5/13 earnings)	-3% ~ -4%	High
Southbound flows turning net seller of Tencent	-2% ~ -3%	High
AI narrative captured by competitors (DeepSeek V4/Qwen3.6/Nightingale 1T)	-1% ~ -2%	Medium
Sector/macro beta (Oil + Geopolitics + Fed Warsh hawkish)	-2% ~ -3%	High
Pre-Q1 earnings risk-off	-1% ~ -2%	Medium
Fundamentals deterioration	0%	Very High (Excluded)

Nature Judgment: ✅ Mixed

70% liquidity/sentiment + 20% long-term AI narrative concern + 10% pre-earnings uncertainty

Key counter-evidence: Duan Yongping sold Tencent puts on