# How the Deep Company Series Generates Comprehensive Research Reports in AI Berkshire

> Discover how the Deep Company Series generates 120,000-word AI Berkshire research suites using an eight-part AI-driven workflow, structured templates, and rigorous fact-checking.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: how-to-guide
- Published: 2026-06-30

---

**The Deep Company Series is an eight-part, AI-driven workflow defined in [`skills/deep-company-series.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/deep-company-series.md) that transforms a single company name into a 120,000-word research suite through structured templates, strict fact-checking protocols, and four execution phases.**

The **deep company series** serves as the primary research engine for the AI Berkshire project, converting a simple company query into a comprehensive, publication-ready investment analysis. Orchestrated through the canonical skill definition in [`skills/deep-company-series.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/deep-company-series.md), this system enforces institutional-grade quality controls across every stage of the writing process.

## Triggering the Research Pipeline

Users initiate the workflow by issuing the slash command `/deep-company-series <company>` within Claude or Codex interfaces. This command launches the "deep-company-series" skill, which acts as the central orchestrator for the entire pipeline.

The skill definition is mirrored in [`codex-skills/deep-company-series/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/deep-company-series/SKILL.md) to ensure identical behavior across environments. Whether running locally or in a cloud agent, the system references the same source file, guaranteeing consistent output formatting and quality standards.

```bash
/deep-company-series 腾讯

```

Local execution loads the YAML skill definition directly from the repository:

```python
python3 -c "import pathlib, yaml; \
skill = yaml.safe_load(open('skills/deep-company-series.md')); \
print(f\"Launching {skill['name']} for {sys.argv[1]}\")" 腾讯

```

## The Eight-Article Template Structure

The skill defines a rigid eight-article template (01-08) that structures the final research suite. Each slot targets specific analytical dimensions with assigned word counts to ensure comprehensive coverage.

The standardized sequence includes:

- **01-认知重置.md** – Cognition reset and foundational context
- **02-护城河.md** – Moat analysis and competitive positioning  
- **03-利润引擎.md** – Profit engine and revenue drivers
- **04-隐性资产.md** – Hidden assets and off-balance-sheet value
- **05-AI时代前景.md** – AI-era outlook and technological disruption
- **06-财务深度.md** – Financial deep-dive and ratio analysis
- **07-管理层评估.md** – Management assessment and governance
- **08-估值与卖出.md** – Valuation metrics and exit signals

Generated articles are stored sequentially under `reports/<公司名>/《看懂<公司名>》/`, creating a navigable research directory:

```

reports/拼多多/《看懂拼多多》/
│   01-认知重置.md
│   02-护城河.md
│   …
│   08-估值与卖出.md
│   00-系列说明.md   ← internal index, not published

```

## Editorial Standards and Quality Gates

Before any content generation begins, the skill enforces strict **writing-style rules** covering tone, vocabulary restrictions, headline formatting, and a blacklist of subjective words. These constraints ensure the output remains data-driven and investment-grade rather than speculative.

The workflow implements two mandatory seven-item checklists:

**The "伪精确" (Pseudo-Precision) Checklist** – Forces verification of weighted probabilities, prevents over-extrapolation, and requires disclosure of holdings or conflicts of interest before saving drafts.

**The "修订时必跑的检查" (Mandatory Revision Checks)** – Ensures cross-article consistency, proper unit labeling for all financial figures, and rigorous source attribution for every statistical claim.

## The Four-Phase Execution Pipeline

The deep company series operates through a structured four-phase workflow defined in the skill file:

**Phase 1 – Research**  
The system ingests five years of annual reports and a minimum of three sell-side analyst analyses. Research agents (`/investment-team` or `/investment-research`) compile foundational data before writing begins.

**Phase 2 – Writing**  
Articles are generated sequentially in the 01-08 order, with each draft stored immediately to `reports/<公司名>/《看懂<公司名>》/0X-XX.md`. This linear approach ensures narrative coherence and logical dependency chains between sections.

**Phase 3 – Cross-Article Scan**  
An "Explore agent" executes automated consistency checks across the entire series. This scan validates numeric consistency, terminology definitions, and internal reference linking using the helper script:

```bash
python3 scripts/run-explore-agent.py \
  --target reports/拼多多/《看懂<公司名>》/ \
  --checks numeric-consistency terminology

```

**Phase 4 – Final Release Check**  
A local `grep` operation scrubs personal identifiers from all files. The system injects the mandatory disclaimer ("不构成投资建议") and verifies no private data remains before marking the series as publication-ready.

## Revision Handling and Compliance

Post-generation modifications follow a structured triage system categorizing feedback as **hard-error**, **subjective**, **granular**, or **unreliable**. Each accepted change triggers a **linked-update scan** that propagates necessary adjustments across all eight articles to maintain internal consistency.

Publication guardrails enforce strict compliance rules: no private data may be exposed, and all outputs must include the standardized investment disclaimer. These checks are hard-coded into the skill definition at lines 84-90 of [`skills/deep-company-series.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/deep-company-series.md), ensuring automated enforcement before any public release.

## Summary

- The **deep company series** uses the skill file [`skills/deep-company-series.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/deep-company-series.md) as its single source of truth for orchestrating AI-generated research.
- It produces **eight standardized articles** (01-08) covering moats, financials, management, and valuation, totaling approximately 120,000 words.
- **Two seven-item checklists** enforce fact-checking and cross-article consistency before publication.
- The **four-phase pipeline** (Research → Writing → Cross-Article Scan → Final Release) ensures institutional-grade quality control.
- Generated content is stored in `reports/<公司名>/《看懂<公司名>》/` with automated scans for numeric consistency and terminology alignment.

## Frequently Asked Questions

### What is the deep company series in AI Berkshire?

The deep company series is a structured AI workflow that converts a single company name into a comprehensive eight-part research report. Defined in [`skills/deep-company-series.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/deep-company-series.md), it coordinates multiple specialized agents to produce approximately 120,000 words of investment-grade analysis covering competitive moats, financial health, and valuation metrics.

### How many articles does the deep company series generate?

The workflow generates **eight distinct markdown articles** numbered 01 through 08, plus an internal index file (00-系列说明.md). The series covers cognition reset, moat analysis, profit engines, hidden assets, AI outlook, financial deep-dives, management assessment, and valuation signals.

### What quality checks ensure accuracy in the deep company series?

The system enforces **two seven-item checklists**: the "伪精确" checklist prevents pseudo-precision and undisclosed conflicts, while the "修订时必跑的检查" ensures cross-article consistency and proper source attribution. Additionally, the "Explore agent" runs automated numeric consistency scans during Phase 3 of the pipeline.

### Where are the generated reports stored?

Completed articles are saved to `reports/<公司名>/《看懂<公司名>》/` within the repository root, following the naming convention `0X-标题.md`. This directory structure separates research outputs by company while maintaining a standardized eight-file format for every analysis.