How to Generate a Deep Company Series Analysis with AI Berkshire

AI Berkshire's /deep-company-series skill orchestrates an eight-article, ~120,000-word research pipeline that produces publication-ready investment analysis through a four-phase multi-agent workflow.

The xbtlin/ai-berkshire repository provides a specialized command-line tool for generating comprehensive deep company series analysis. This skill leverages a multi-agent framework to automate the creation of rigorous, fact-checked investment research suitable for professional publication. When you generate a deep company series analysis, the system produces eight interconnected markdown articles stored in a structured directory hierarchy.

Understanding the /deep-company-series Skill Architecture

The deep company series functionality is defined in skills/deep-company-series.md, which serves as the orchestration blueprint for the entire research pipeline. This skill implements the same four-master framework used across AI Berkshire's investment research tools, ensuring consistency with rigorous financial analysis standards.

The Eight-Article Blueprint

The skill automatically generates a standardized series structure beginning with a 00-系列说明.md index page followed by eight specialized articles (01-08). According to the skill definition in skills/deep-company-series.md (lines 20-32), each article targets specific focus questions with defined word-count ranges:

• Article 01: Foundational misconceptions and business model clarity
• Articles 02-07: Sequential deep-dives into financial metrics, competitive positioning, management quality, and risk factors
• Article 08: Valuation synthesis with entry/exit signals

Writing Discipline and Quality Controls

The skill enforces strict editorial standards through two enforcement mechanisms defined in skills/deep-company-series.md (lines 47-63). First, a "禁用词" (banned words) list eliminates subjective phrasing and promotional language. Second, the "写作风格规范" (writing style specifications) mandate neutral tone and bullet-point structure for scannability.

Before publication, a 7-point "严苛事实核查 Checklist" (Strict Fact-Check Checklist) runs validation across all drafts (lines 66-86). This validates cross-article number consistency, data-source labeling, and terminology standardization.

The Four-Phase Research Workflow

When you execute /deep-company-series, the skill initiates a structured workflow defined in skills/deep-company-series.md (lines 98-130).

Phase 1 – Research and Data Collection

The system requires three input categories to seed the series:

• Annual and quarterly reports (10-K, 10-Q equivalents)
• Minimum three independent sell-side research reports
• Optional pre-computed analysis from /investment-team or /investment-research skills

This data ingestion phase establishes the factual foundation before any drafting begins.

Phase 2 – Sequential Drafting

The skill writes articles sequentially (01→08) into reports/{Company}/《看懂{Company}》/ directories. As specified in skills/deep-company-series.md (lines 107-112), drafts remain in local storage without immediate Git commits, allowing for iterative refinement between articles.

Phase 3 – Cross-Article Consistency Scan

After all eight drafts complete, an Explore agent validates consistency across the series. This scan checks for numeric value alignment, standardized terminology usage, and cross-reference accuracy between articles (lines 14-21).

Phase 4 – Privacy Scan and Publication

Before final commit, a grep-based privacy scan executes to detect potential PII leaks:

grep -r "linxuan\|/Users/" reports/{Company}/《看懂{Company}》/

If no matches return, the series passes validation and becomes eligible for Git commit and push operations (lines 24-30).

Step-by-Step Implementation Guide

To generate a deep company series analysis in your local environment, follow these installation and execution steps.

Installation

First, install the skill using the provided installation script:

# Clone the repository first

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

# Install the skill (choose based on your AI coding tool)

./scripts/install-codex-skills.sh

# OR

./scripts/install-claude-commands.sh

Running the Series Generator

Execute the skill with a target company name. For example, to analyze PDD Holdings:

/deep-company-series 拼多多

The system responds with creation confirmation and a file list:

✅ 已创建 “拼多多” 系列项目
📄 生成 00‑系列说明.md、01‑你以为你看懂了拼多多，其实没有.md … 08‑多少钱值得买，什么信号必须卖.md
🔎 请提供公司年报、最近季报以及至少 3 份卖方研报链接

Reviewing Drafts

During the drafting phase, inspect individual articles before final publication:

cat reports/拼多多/《看懂拼多多》/01-你以为你看懂了拼多多，其实没有.md

Final Verification and Commit

After the cross-article scan completes, run the privacy verification:

grep -r "linxuan\|/Users/" reports/拼多多/《看懂拼多多》/

If no privacy warnings appear, commit the series:

git add reports/拼多多/《看懂拼多多》/
git commit -m "Add deep-company-series for 拼多多"
git push

Key Source Files and Dependencies

The deep company series analysis relies on several critical components within the AI Berkshire repository:

• skills/deep-company-series.md: Core skill definition containing the eight-article templates, writing rules, and workflow orchestration logic.
• tools/financial_rigor.py: Supporting utility for precise financial calculations, including market-cap verification and cross-validation functions used throughout the series generation.
• reports/腾讯/《看懂腾讯》/: Reference implementation showing the complete output structure for Tencent (腾讯), demonstrating the expected directory layout and article formatting.

Summary

• AI Berkshire provides the /deep-company-series skill for automated generation of comprehensive eight-article investment research series.
• The workflow follows four phases: Research → Draft → Cross-Article Scan → Publication, enforced through skills/deep-company-series.md.
• Each series produces approximately 120,000 words across eight standardized articles with strict fact-checking requirements.
• Output files generate in reports/{Company}/《看懂{Company}》/ with mandatory privacy scanning before Git commit.
• The system requires annual reports, quarterly filings, and minimum three sell-side research reports as input seeds.

Frequently Asked Questions

What input data does the deep company series require?

The skill requires annual reports, recent quarterly filings, and at least three independent sell-side research reports to initialize the research phase. You may optionally supplement this with existing outputs from /investment-team or /investment-research skills to accelerate the process.

How long does it take to generate a complete eight-article series?

While timing varies by company complexity, the skill is designed to produce approximately 120,000 words across eight articles. The sequential drafting process (Phase 2) runsarticle-by-article to maintain narrative consistency, with the cross-article scan (Phase 3) running only after all drafts complete.

Can I customize the article templates for specific industries?

The current implementation in skills/deep-company-series.md uses fixed eight-article templates (01-08) with specific focus questions. While the titles and focus areas are standardized for investment analysis consistency, you can modify the underlying skill file before installation to adjust word-count targets or specific focus questions for sector-specific needs.

How does the system ensure factual accuracy across all eight articles?

The 7-point Strict Fact-Check Checklist runs automated validation for numerical consistency, data-source labeling, and terminology standardization. Additionally, the Cross-Article Scan (Phase 3) employs an Explore agent to verify that metrics referenced in early articles match values cited in later valuation synthesis sections.

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