# How the AI Berkshire Investment Team Skill Simulates a Professional Investment Research Team

> Discover how the AI Berkshire investment team skill simulates professional research using a nine-stage multi-agent workflow, specialized roles, master investor models, and automated validation to deliver institutional-grade ana...

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: how-to-guide
- Published: 2026-07-06

---

**The AI Berkshire investment team skill orchestrates a nine-stage multi-agent workflow that mirrors institutional research teams by assigning specialized analyst roles, embedding master-investor mental models, and enforcing rigorous data validation through automated financial tooling and mandatory audit gates.**

The `investment-team` skill in the xbtlin/ai-berkshire repository transforms a single company query into a comprehensive, parallelized research operation. By leveraging the **AI Berkshire investment team skill**, the system autonomously creates a virtual team of analyst agents, assigns master-investor frameworks, and enforces institutional-grade data verification. This approach faithfully replicates the division of labor and rigorous compliance standards found in professional investment research houses.

## The 9-Stage Investment Research Pipeline

The simulation follows a strict pipeline defined in [`skills/investment-team.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-team.md), progressing from team definition to final audit. Each stage maps to real-world research workflows found in institutional asset management firms.

### Stage 1: Team Definition and Framework Selection

The skill establishes a five-role analytical framework comprising the **team-lead**, **business-analyst**, **financial-analyst**, **industry-researcher**, and **risk-assessor**. Each analyst role maps to a specific master-investor's perspective—段永平 (Duan Yongping), 巴菲特 (Buffett), 芒格 (Munger), and 李录 (Li Lu)—ensuring multi-dimensional coverage of the investment thesis.

### Stage 2: Information Richness Assessment

Before data collection begins, the system evaluates the target company's "information richness" (信息丰富度) and assigns an A/B/C rating. This determines analysis depth, with A-level targets receiving full quantitative rigor while restricting speculative conclusions for data-scarce entities.

### Stage 3: Team Creation

The skill invokes the `TeamCreate` tool to instantiate a named team (e.g., `meituan-research`) with the `team-lead` agent type. This establishes the orchestration layer that will manage the parallel workflow.

### Stage 4: Parallel Task Generation

Four concurrent `TaskCreate` jobs are generated, one per analyst role (excluding the team-lead). Each task includes a detailed subject, description, and `activeForm` that embeds the assigned master-investor's analytical framework and strategic questions.

### Stage 5: Concurrent Agent Execution

All four tasks are dispatched in a single message with `run_in_background = true`, launching **general-purpose** sub-agents simultaneously. This parallelism mirrors the real-world efficiency of research teams working asynchronously on different aspects of the same coverage universe.

### Stage 6: Data Gathering and Financial Validation

Each sub-agent executes `WebSearch` calls to retrieve filings, news, and market data. Critically, all quantitative data passes through the **Financial-Rigor Toolkit** ([`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)) for exact-decimal verification of market capitalization, valuation multiples, and cross-source consistency.

### Stage 7: Progress Monitoring

The team-lead aggregates real-time progress updates, displaying which agents have completed their analysis and highlighting key takeaways from each stream. This simulates the stand-up meetings and status updates common in research departments.

### Stage 8: Synthesis and Master Report Compilation

Upon completion of all four reports, the team-lead issues a `shutdown_request` to terminate the agents, then compiles a master report. This document fuses the four perspectives, adds a one-sentence investment conclusion, generates a rating table, and includes a detailed checklist.

### Stage 9: Data Audit and Compliance Verification

The final markdown file undergoes the "准出" (approval/exit) flow through [`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py). This extracts a random sample of data points, re-fetches them from defined sources, and verifies that every figure is within **1%** of the source value before publication.

## Financial Rigor and Data Validation

Unlike standard LLM analysis prone to arithmetic hallucinations, the **AI Berkshire investment team skill** integrates the [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) toolkit for exact decimal arithmetic. This command-line utility performs verification of market capitalization, valuation multiples, and cross-source consistency using Python's `Decimal` type to eliminate floating-point errors.

```bash
python3 tools/financial_rigor.py verify-market-cap \
  --price 510 --shares 9.11e9 --reported 4.65e12 --currency HKD

```

The tool outputs a detailed breakdown, flags any deviation exceeding 5%, and returns a Boolean that the skill inserts into the final report. This eliminates "LLM mental-math" errors that plague ad-hoc analysis.

Additionally, the toolkit supports **Benford's Law** checks, three-scenario valuation modeling, and dual-source verification policies defined in [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md) for US, HK, and A-share markets.

## The Audit and Compliance Gate

Before any report reaches the user, it must pass the **data-audit (准出) flow** implemented in [`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py). This secondary verification process performs three critical functions:

- Extracts a random sample of quantitative data points from the generated markdown
- Re-fetches the underlying data from primary sources (macrotrends, stockanalysis, aastocks, 东方财富, 巨潮资讯)
- Verifies that every figure is within **1%** of the source value

This compliance layer ensures the **AI Berkshire investment team skill** meets institutional standards for data accuracy, comparable to the review processes at professional research houses.

## Key Implementation Files

The simulation relies on these specific components within the xbtlin/ai-berkshire repository:

- **[`skills/investment-team.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-team.md)** — Master prompt defining the team architecture, role-specific prompts, and orchestration flow via `TeamCreate` and `TaskCreate` tools.
- **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)** — Command-line toolkit for exact decimal verification, Benford checks, and three-scenario valuation analysis.
- **[`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py)** — Implements the post-generation "准出" workflow for secondary data verification and quality gates.
- **[`skills/investment-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-research.md)** — High-level framework specifying data-source rules, audit steps, and research methodologies that invoke the team skill.
- **[`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md)** — Defines dual-source collection policies for US, HK, and A-share markets.

## Summary

The **AI Berkshire investment team skill** replicates professional investment research through these key mechanisms:

- **Division of labor** across four specialized analytical roles (business, financial, industry, risk), each adopting the mental models of master investors like Buffett and Munger.
- **Parallel execution** via concurrent `TaskCreate` calls that minimize latency while maintaining specialized focus.
- **Rigorous data validation** using exact `Decimal` arithmetic in [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) to eliminate computational errors.
- **Compliance auditing** through the [`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py) "准出" gate, enforcing ≤1% deviation tolerance before publication.
- **Master-investor frameworks** embedded in prompt templates to ensure multi-dimensional coverage of business quality, moats, management, and valuation.

## Frequently Asked Questions

### How does the AI Berkshire investment team skill handle data accuracy?

The skill enforces a two-layer validation system. First, [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) verifies all calculations using exact decimal arithmetic and cross-source consistency checks during analysis. Second, [`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py) performs a post-generation audit that re-fetches a random sample of data points to verify they are within 1% of source values before the report is delivered.

### What are the four analytical roles in the investment team simulation?

The simulation creates four specialized analyst agents: the **business-analyst** (focusing on business quality and moats), the **financial-analyst** (examining financial statements and valuation), the **industry-researcher** (assessing competitive positioning and trends), and the **risk-assessor** (identifying downside scenarios). A fifth **team-lead** agent orchestrates the workflow and synthesizes findings without executing a parallel task.

### Which master investors' frameworks does the skill utilize?

The prompt templates embed the analytical frameworks of four legendary investors: **Warren Buffett** (business moats and management quality), **Charlie Munger** (risk assessment and mental models), **Duan Yongping** (business quality and network effects), and **Li Lu** (civilization-level trends and industry structure).

### How does the parallel execution work technically?

The skill generates four `TaskCreate` calls with `run_in_background = true` and dispatches them in a single message. This launches four general-purpose sub-agents concurrently, each executing web searches and analysis independently while the team-lead monitors progress. This architecture mimics how human research teams work asynchronously on different aspects of the same coverage stock.