# AI-Berkshire Use Cases: Multi-Agent Value Investing Framework Explained

> Explore AI-Berkshire use cases for AI-driven value investing. This framework uses four AI agents for deep fundamental analysis, portfolio monitoring, and financial data validation.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: deep-dive
- Published: 2026-07-08

---

**AI-Berkshire is a modular, AI-driven research framework that orchestrates four value-investment masters (巴菲特, 芒格, 段永平, 李录) as independent agents to perform deep fundamental analysis, portfolio monitoring, and financial data validation.**

The open-source **ai-berkshire** project implements a three-layer architecture that transforms traditional investment research into parallel, multi-agent workflows. By invoking specific skills, users can trigger automated research pipelines that leverage specialized tools for financial rigor and cross-source validation.

## Core Architecture: The Three-Layer Framework

Before diving into specific use cases, understanding the architecture explains how the system scales from quick screens to deep-dive analysis.

### Skill Layer

The **skill layer** defines *what* you want to accomplish. Located in `skills/*.md`, this catalogue contains 19 distinct skills ranging from `/investment-research` to `/news-pulse`. Each skill declaration specifies the orchestration logic, required data inputs, and expected output format.

### Agent Layer

When you invoke a skill, the system spawns **four parallel agents**—one for each investment master. According to the architecture documented in [`AGENTS.md`](https://github.com/xbtlin/ai-berkshire/blob/main/AGENTS.md), these agents independently search, evaluate, and challenge one another's conclusions. A team-lead agent aggregates the results, creating a consensus view that reduces individual bias while preserving dissenting perspectives.

### Tool Layer

The **tool layer** ensures numerical precision and data integrity. Key files like [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) and [`tools/stock_screener.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/stock_screener.py) perform exact decimal calculations, cross-source validation, and data scraping. This layer guarantees that agent discussions are grounded in verified financial data rather than hallucinated figures.

## AI-Berkshire Use Cases: Skill-by-Skill Breakdown

The following use cases map directly to the 19-skill catalogue and demonstrate how the framework handles different research intensities.

### Full-Company Deep Dives

**`/investment-research`** triggers a comprehensive four-master analysis of any listed company. The system generates a structured decision memo scoring business quality, competitive moat, management integrity, and safety margin. This is the most thorough use case, typically used for target conviction building.

**`/investment-team`** delivers the same analytical depth as the deep dive but executes in parallel, producing quicker results. The output includes a consolidated multi-agent report with a one-sentence verdict and detailed scoring tables, making it ideal for time-sensitive opportunities.

### Management and Governance Analysis

**`/management-deep-dive`** focuses exclusively on leadership evaluation. This skill assembles narrative assessments of CEO strengths/weaknesses, governance structures, and corporate culture by cross-referencing multiple sources through the four agents' distinct perspectives.

### Private and Unlisted Company Research

**`/private-company-research`** addresses the challenge of fragmented data on unlisted firms. The agents assemble public records, trademark filings, and supply chain data to construct a company portrait with confidence-tagged data points and valuation ranges where traditional screeners fail.

### Earnings and Event Analysis

**`/earnings-review`** performs line-by-line interpretation of earnings filings. The tool extracts full-text data, calculates key ratio changes, and generates commentary on whether results support or contradict existing investment theses.

**`/news-pulse`** provides rapid attribution of price swings. Within minutes, the agents analyze whether a sudden drop (e.g., "拼多多 跌12% 一周内") stems from fund flows, sentiment shifts, or fundamental deterioration, delivering a short narrative with high confidence.

### Screening and Funnel Analysis

**`/industry-funnel`** and **`/industry-research`** implement a two-stage screening process. The funnel skill scans entire markets to shortlist ≤10 candidates, while the research skill produces sector-level investment theses. This supports top-down portfolio construction.

**`/quality-screen`** and **`/investment-checklist`** offer rapid pre-screening. The six-gate Buffett-style checklist (`/investment-checklist`) returns pass/fail verdicts per gate plus a "mirror test" conclusion, allowing investors to kill bad ideas quickly before committing resources to deep research.

### Portfolio Monitoring and Thesis Tracking

**`/portfolio-review`**, **`/thesis-tracker`**, and **`/thesis-drift`** provide ongoing surveillance of existing holdings. These skills generate allocation suggestions, thesis-status dashboards, and drift analysis alerts when company fundamentals diverge from original investment cases.

### Specialized Perspective Queries

**`/dyp-ask`** (段永平 Ask) functions as a thought-experiment tool. Rather than analyzing a specific security, this skill frames problems from 段永平's philosophical perspective, returning answers prefaced with "段永平 would ask..."

### Financial Data Validation

**`/financial-data`** triggers [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) to retrieve and cross-validate key metrics like market cap and EPS. The output includes tabulated data with source citations and variance alerts when databases disagree, ensuring agents work with accurate inputs.

## How to Invoke Skills: Command Examples

The repository supports both Claude Code and Codex client interfaces. The underlying logic remains identical—both trigger the skill scripts that automatically spin up the four master agents, pull data via [`tools/xueqiu_scraper.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/xueqiu_scraper.py) or [`tools/ashare_data.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/ashare_data.py), validate numbers with [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py), and assemble markdown reports.

### Claude Code (npm-installed)

```bash

# Deep research on 腾讯

/investment-research 腾讯

# Parallel team research on 美团

/investment-team 美团

# Quick safety-margin filter on a list of stocks

/investment-checklist 腾讯, 阿里巴巴, 美团, 拼多多

# Earnings review of PDD for 2025 Q4

/earnings-review PDD 2025Q4

# Industry funnel on AI算力

/industry-funnel AI算力

# News-pulse attribution for a sudden drop

/news-pulse 拼多多 跌12% 一周内

```

### Codex (after [`install-codex-skills.sh`](https://github.com/xbtlin/ai-berkshire/blob/main/install-codex-skills.sh))

```bash

# Full company analysis (Codex syntax)

investment-research 腾讯

# Fast parallel research

investment-team 美团

# Run the six-gate checklist

investment-checklist 腾讯, 阿里巴巴, 美团, 拼多多

# Earnings deep-read

earnings-review PDD 2025Q4

# Industry screening

industry-funnel AI算力

# Quick news attribution

news-pulse 拼多多 跌12% 一周内

```

## Key Implementation Files

Understanding these source files helps advanced users customize workflows:

- **`skills/*.md`** – Canonical skill definitions for Claude Code source
- **`codex-skills/*/SKILL.md`** – Auto-generated Codex skill packages (generated via [`scripts/sync-codex-skills.py`](https://github.com/xbtlin/ai-berkshire/blob/main/scripts/sync-codex-skills.py))
- **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)** – Decimal-based valuation and cross-validation utilities
- **[`tools/stock_screener.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/stock_screener.py)** – Fast screening logic leveraged by `/quality-screen`
- **[`AGENTS.md`](https://github.com/xbtlin/ai-berkshire/blob/main/AGENTS.md)** – Design document describing the three-layer architecture
- **`reports/`** – Example output reports produced by the skills

## Summary

- **AI-Berkshire** coordinates four value-investment masters as parallel agents to eliminate single-point-of-view bias in financial analysis.
- The **19-skill catalogue** covers the full research spectrum from quick checklist screening (`/investment-checklist`) to comprehensive due diligence (`/investment-research`).
- **Architecture consistency** means every skill automatically engages the tool layer ([`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)) for data validation and the agent layer for multi-perspective analysis.
- **Flexible invocation** supports both Claude Code (slash commands) and Codex (natural language) interfaces with identical underlying logic.
- **Use cases span the investment lifecycle**: idea generation (screening), conviction building (deep dives), position monitoring (thesis tracking), and event reaction (news-pulse).

## Frequently Asked Questions

### What is the difference between `/investment-research` and `/investment-team`?

Both skills perform comprehensive four-master analysis, but `/investment-team` executes agent tasks in parallel rather than sequentially, delivering faster results for time-sensitive opportunities. The output format differs slightly—`/investment-team` emphasizes a one-sentence verdict upfront, while `/investment-research` provides a more detailed narrative structure.

### Can AI-Berkshire analyze private companies or stocks not listed on major exchanges?

Yes. The **`/private-company-research`** skill specifically handles unlisted entities by aggregating fragmented public data from trademark databases, supply chain records, and local business registries. The tool layer assigns confidence scores to each data point, acknowledging higher uncertainty while still enabling fundamental analysis.

### How does the framework ensure financial data accuracy and prevent AI hallucination?

The **tool layer** enforces financial rigor through [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py), which performs exact decimal calculations and cross-source validation. When agents retrieve metrics like EPS or market cap, the system queries multiple data providers ([`tools/xueqiu_scraper.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/xueqiu_scraper.py), [`tools/ashare_data.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/ashare_data.py)) and flags variances above defined thresholds, ensuring all subsequent analysis uses verified numbers.

### Is it possible to chain multiple skills together for a complete workflow?

Yes. Users commonly chain skills to balance depth versus cost. For example, run `/quality-screen` or `/investment-checklist` on a universe of stocks to filter candidates, then promote survivors to `/investment-research` for full deep-dive treatment. The markdown output format from each skill is designed to pipe into downstream tools or serve as context for subsequent invocations.