How to Implement the Four-Master Investment Framework (Buffett · Munger · Duan · Li) in AI Berkshire

The AI Berkshire repository implements a staged, multi-agent pipeline where Duan Yong-ping analyzes business essence, Warren Buffett evaluates economic moats, Charlie Munger applies inversion thinking, and Li Lu assesses civilizational trends, converging into a unified decision matrix that outputs price-anchored buy/hold/sell recommendations.

The xbtlin/ai-berkshire open-source project codifies the value-investment mental models of these four masters into a reproducible AI workflow. This four-master investment framework structures research as a sequential pipeline—visualized in assets/architecture-en.svg and documented in README_EN.md—where each analytical layer feeds into the next, ensuring outputs pass rigorous pass/fail/gray-zone criteria before generating final recommendations.

Pipeline Architecture and Data Flow

The framework processes raw inputs through seven distinct stages orchestrated by the investment-research skill defined in skills/investment-research.md. Data collection begins with the financial-data skill (skills/financial-data.md), which gathers annual reports, 13-F filings, and macro data, cross-validating them into JSON storage layers (data/fundamentals.json, data/watchlist.json) for downstream consumption.

Stage 1: Business Essence (Duan Yong-ping)

Duan’s analytical lens extracts the business core: product-market fit, competitive advantage mechanics, and founder-led capital allocation. Unlike ratio-based screening, this stage emphasizes how the business creates value. The investment-research skill explicitly implements the "Business Essence" block to evaluate these qualitative factors before quantitative analysis begins.

Stage 2: Moat Durability (Warren Buffett)

Buffett’s layer evaluates competitive fortress durability through network effects, switching costs, economies of scale, and cash-machine quality. The investment-checklist skill, defined in skills/investment-checklist.md, encodes Buffett’s six-gate pre-buy checklist. Final reports surface this analysis in structured ratings such as "Moat (Buffett) | Wide and widening".

Stage 3: Inversion Analysis (Charlie Munger)

Munger’s inversion principle systematically asks, "What could cause this investment to fail?" The investment-research skill contains a dedicated "Inversion (Munger)" stage that identifies downside catalysts, regulatory risks, and management misalignment. This stage produces a risk-inversion score that feeds directly into the final decision matrix.

Stage 4: Management Assessment (Duan + Buffett)

This joint evaluation layer combines both masters’ emphasis on people. It assesses CEO quality, capital allocation discipline, and interest alignment between owners and management. The assessment appears in final outputs as the "Management (Duan + Buffett)" rating, requiring both qualitative judgment and quantitative insider-ownership metrics.

Stage 5: Civilizational Trends (Li Lu)

Li Lu contributes a macro-civilizational perspective evaluating long-term industry secular trends, geopolitical stability, and existential durability. This stage determines whether the company will still exist in ten years. The investment-research skill captures this view in the "Civilizational Trends (Li Lu)" column, filtering out businesses facing technological obsolescence or regulatory extinction.

Decision Matrix and Memo Generation

All five analytical scores converge through the investment-memo-craft skill defined in codex-skills/investment-memo-craft/SKILL.md. This Codex-only skill formats the final research memo, ensures consistent typography, and applies the four-master decision matrix logic to yield clear Buy / Hold / Sell recommendations with specific price bands. The matrix requires that all five lenses align (or explicitly flag gray-zone exceptions) before issuing a buy signal.

Parallel Agent Execution

The framework leverages a multi-agent architecture described in the investment-team skill (codex-skills/investment-team/SKILL.md). Four specialized agents—Buffett, Munger, Duan, and Li—run in parallel, each invoking its respective sub-skill, then synchronize findings through the investment-research skill’s orchestration layer. This parallelization maintains analytical independence while ensuring convergent conclusions.

CLI Commands for Implementation

Run the complete pipeline or isolate specific master analyses using the ai-berkshire CLI wrapper:

# Execute the full four-master research pipeline on a ticker

ai-berkshire /investment-research AAPL

# Isolate Buffett’s moat analysis using the six-gate checklist

ai-berkshire /investment-checklist AAPL --stage moat

# Generate Li Lu’s 10-year civilizational outlook

ai-berkshire /investment-research AAPL --stage li-trends

# Compile all analyses into a formatted investment memo

ai-berkshire /investment-memo-craft AAPL --format markdown

Each command invokes the respective Codex skills and returns Markdown blocks suitable for direct insertion into research reports.

Summary

• The four-master investment framework processes data through seven sequential stages from collection to decision, as defined in skills/investment-research.md.
• Each master contributes a distinct analytical lens: Duan (business essence), Buffett (moat), Munger (inversion), and Li (civilizational trends).
• The investment-checklist skill encodes Buffett’s six-gate pre-buy criteria, while the investment-memo-craft skill generates final price-anchored signals.
• Parallel agent execution via codex-skills/investment-team/SKILL.md allows simultaneous processing by four specialized agents before synchronization.

Frequently Asked Questions

Which source files define the core four-master pipeline?

The primary orchestration logic resides in skills/investment-research.md and codex-skills/investment-research/SKILL.md. Buffett’s specific six-gate checklist is implemented in skills/investment-checklist.md, while the multi-agent coordination is defined in codex-skills/investment-team/SKILL.md.

How does the framework handle data validation and storage?

Raw inputs are collected by the financial-data skill and cross-validated before persistence in data/fundamentals.json and data/watchlist.json. This JSON data layer ensures consistent formatting for downstream analytical stages.

Can I run individual master analyses without the full pipeline?

Yes. Use stage-specific flags such as --stage moat for Buffett’s analysis or --stage li-trends for Li Lu’s civilizational outlook. The full integrated pipeline runs via /investment-research without flags.

What determines the final buy/hold/sell recommendation?

The investment-memo-craft skill converges five quantitative/qualitative scores—Business Essence, Moat, Inversion, Management, and Civilizational Trends—into a weighted decision matrix. This matrix yields a definitive signal only when the majority of lenses align, with explicit price bands derived from DCF and comparative valuation models.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}