How to Use the Investment-Research Skill in AI Berkshire: A Complete Guide

The /investment-research command in AI Berkshire executes a seven-step master-master workflow defined in skills/investment-research.md to generate audited valuation reports for any public company.

AI Berkshire is an open-source investment research framework that automates institutional-grade equity analysis. The investment-research skill orchestrates data collection, cross-validation, and multi-scenario valuation through a declarative markdown workflow and deterministic Python tools, producing reproducible research reports every time you run the command.

Understanding the Three-Layer Architecture

Skill Layer

The skill layer defines what the command does. The workflow is purely declarative, residing in skills/investment-research.md. This file specifies a seven-module analysis pipeline: data collection, business model analysis, moat evaluation, reverse-thinking exercises, management assessment, civilization trend analysis, and final valuation.

Agent Layer

When you invoke /investment-research <Company>, the agent layer executes the skill sequentially. It spawns background tasks to gather data, then walks through each analytical module, appending "master's 追问" (follow-up questions) at each stage to enforce rigorous reasoning.

Tool Layer

Deterministic calculations are delegated to the tool layer. The skill relies on tools/financial_rigor.py for exact-decimal operations including market-cap verification (verify-market-cap), cross-source validation (cross-validate), and three-scenario valuation modeling.

Step-by-Step Execution Flow

The command follows a twelve-stage pipeline from input to audited output:

1. Input Parsing: You invoke /investment-research 腾讯 (or any ticker/company name) via the CLI front-end.

2. Data Collection: The agent spawns a background Task scraping financial statements and macro-trends, using the source list defined in skills/financial-data.md.

3. Cross-Validation: Key numbers are passed to tools/financial_rigor.py for market-cap verification and multi-source cross-validation to ensure accuracy.

4. Analysis Modules: The agent sequentially executes the seven master-master modules (business model, moat, reverse-thinking, management, civilization trend, valuation), building a Markdown section for each.

5. Report Assembly: The completed report is written to ~/[Company]投资研究报告.md.

6. Audit Verification: tools/report_audit.py extracts a random 15% sample of data points for a final sanity check before marking the report as准出 (approved for release).

Installation and Setup

To enable the skill in your environment, first clone the repository:

git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire

For Codex users, install the skills and optional prompt adapters:

./scripts/install-codex-skills.sh
./scripts/install-codex-prompts.sh

For Claude Code users, run the dedicated installer:

./scripts/install-claude-commands.sh

Restart your Codex or Claude client to load the new skill definitions from codex-skills/investment-research/SKILL.md.

Running the Investment-Research Command

After installation, invoke the skill with a company name or ticker:

Codex CLI:

investment-research 阿里巴巴

Claude Code (slash command):

/investment-research 拼多多

Each command automatically:

• Fetches financial data from the sources listed in skills/financial-data.md
• Validates numbers using tools/financial_rigor.py
• Generates a Markdown report at ~/[Company]投资研究报告.md
• Runs a 15% random audit via tools/report_audit.py

Key Files and Their Roles

• skills/investment-research.md: The master markdown defining the seven-module workflow, data-source table, and required tool calls.
• codex-skills/investment-research/SKILL.md: Generated Codex wrapper enabling CLI invocation.
• tools/financial_rigor.py: Python utilities for market-cap verification, multi-source cross-validation, and three-scenario valuation.
• tools/report_audit.py: Extracts random data samples from generated reports for final verification.
• scripts/install-codex-skills.sh: Helper script copying skill definitions into ~/.codex/skills.
• skills/financial-data.md: Reference list of approved financial data sources.

Summary

• The investment-research skill uses a declarative markdown workflow in skills/investment-research.md to define a seven-step research process.
• Execution relies on a three-layer architecture: Skill (declaration), Agent (orchestration), and Tool (deterministic calculation).
• Validation tools (tools/financial_rigor.py) ensure data accuracy through market-cap verification and cross-source validation.
• Audit requirements (tools/report_audit.py) mandate a 15% random sample check before report approval.
• Reports are saved to ~/[Company]投资研究报告.md and require no manual formatting.

Frequently Asked Questions

What is the seven-step workflow in the investment-research skill?

The workflow defined in skills/investment-research.md comprises: data collection, business model analysis, moat evaluation, reverse-thinking exercises, management assessment, civilization trend analysis, and final valuation. Each module appends a "master's 追问" (follow-up question) to enforce analytical rigor.

How does the skill validate financial data?

The skill delegates validation to tools/financial_rigor.py, which performs market-cap verification via verify-market-cap and cross-source checks via cross-validate. These deterministic Python functions ensure exact-decimal calculations and reproducible results across multiple data sources.

Where are the generated reports saved?

Completed research reports are written to ~/[Company]投资研究报告.md in your home directory. Before finalization, tools/report_audit.py performs a mandatory 15% random sample audit to verify data integrity against original sources.

Can I use the investment-research skill with Claude Code instead of Codex?

Yes. Run ./scripts/install-claude-commands.sh to install the slash-command adapter, then invoke the skill using /investment-research <Company>. The underlying workflow and validation logic remain identical across both clients.

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