How to Configure Deep-Company-Series for 8-Article Long-Form Research Output

To configure the deep-company-series skill for maximum 8-article output, select the high-complexity path in skills/deep-company-series.md, allocate all eight core research pillars, and set word-count targets ranging from 4,000 to 12,000 characters per article.

The deep-company-series skill in the xbtlin/ai-berkshire repository generates multi-article research series that provide textbook-level analysis of single companies. When configured for the maximum 8-article format, it produces comprehensive coverage suitable for conglomerates with multiple business lines and complex asset structures like Tencent.

Select the High-Complexity Path

The 8-article configuration is explicitly designed for companies with intricate operational profiles. According to the complexity table defined in skills/deep-company-series.md (lines 26-30), this tier accommodates firms with multiple business lines, hidden investment portfolios, and extensive management-team history.

High-complexity indicators include:

• Diverse revenue streams across unrelated sectors
• Significant off-balance-sheet assets or holding-company structures
• Historical M&A activity requiring deep institutional memory

Allocate the Eight Core Research Pillars

The skill defines eight "主轴" (core pillars) that each become a standalone article. For the full 8-article series, implement the complete template from lines 42-53 of skills/deep-company-series.md:

1. "你以为你看懂了 X，其实没有" — Cognition reset that challenges superficial assumptions
2. "X 的护城河" — Moat analysis covering competitive advantages
3. "X 的最大利润引擎" — Profit engine identification and unit economics
4. "X 藏在账上的另一家公司" — Hidden assets and investment portfolio deep-dive
5. "AI/时代变量" — AI and era-specific variable impact assessment
6. "用巴菲特方式拆 X 的财报" — Financial statement analysis using Buffett methodology
7. "管理层金句" — Management team evaluation and historical decision analysis
8. "多少钱值得买，什么信号必须卖" — Valuation framework and exit criteria

Each pillar maps to a separate markdown file in the final output.

Configure Article Length Targets

The skill specifies precise word-count ranges to maintain consistent depth across the series. Set your prompt parameters to target these ranges as documented in lines 46-53:

• Article 1 (Intro): 4,000–5,000 characters
• Articles 2–7 (Body): 6,000–8,000 characters each
• Article 8 (Valuation): 10,000–12,000 characters

These targets ensure the final decision article receives appropriate analytical weight while maintaining reader engagement throughout the series.

Set Up the Series Infrastructure

Create the Series Index File

Before generating content, create a 00-系列说明.md file that serves as the internal table of contents. This file lists all eight articles in sequence but remains unpublished (lines 55-56). The index ensures logical flow between the cognition reset opening and the valuation conclusion.

Establish the Output Directory Structure

Persist each article under the path reports/{公司名}/《看懂{公司名}》-{YYYYMMDD}/0X-XX.md. The "目录冲突规则" (directory conflict rules) in lines 58-60 mandate that if a previous series exists for the same company, you must create a new dated folder rather than overwriting existing files.

Structure Individual Articles

Every article in the series must follow the standardized skeleton defined in lines 65-68 of the skill file.

Required Header Block

Each file must begin with a header block containing the series name and article number. This metadata enables the cross-article validation tools to identify and link related files.

Content Skeleton Requirements

• Opening Hook: Begin with a "数字/曲线/惊人事实" (number/chart/surprising fact) hook in the first paragraph (lines 66-68)
• Mid-Article Structure: Develop the pillar-specific analysis using the prescribed frameworks
• Closing Elements: Include a "本篇要点回顾" (key points review) bullet list and a "下期预告" (next episode teaser) at the end (lines 67-69)

Enforce Quality Controls

Run the 7-Item Fact-Check Checklist

Before finalizing each article, execute the checklist defined in lines 24-34. This verification ensures:

• Cross-article number consistency
• Standardized term definitions
• Elimination of duplicate counting across pillars
• Source attribution accuracy

Execute Cross-Article Validation

After drafting all eight articles, launch the Explore agent to automatically verify consistency across the series. This agent checks for numerical alignment, terminology standardization, and cross-reference integrity (lines 63-70).

Style and Privacy Requirements

Adhere to the style guide in lines 87-94: maintain a direct, data-first tone; avoid absolute-value words; and ensure the first 18-20 characters function as a "stand-alone" mobile preview.

Before committing, perform a grep search for personal identifiers. Push the series to the repository only after explicit user approval (lines 71-78).

Programmatic Configuration

Invoke the skill via the AI-Berkshire CLI to generate the configured 8-article series:

from pathlib import Path
import subprocess

def generate_deep_series(company_name: str, complexity: str = "high"):
    """
    Wraps the deep-company-series skill to produce an 8-article series.
    """
    prompt = f"""# {company_name}

    $ARGUMENTS = {company_name}
    """
    result = subprocess.run(
        ["ai-berkshire", "run", "deep-company-series"],
        input=prompt.encode(),
        capture_output=True,
    )
    return result.stdout.decode()

# Generate for a high-complexity firm

print(generate_deep_series("腾讯"))

Alternatively, use the bash interface:

ai-berkshire run deep-company-series --company "腾讯" --complexity high

Both methods reference the master template at skills/deep-company-series.md and output eight markdown files following the architectural specifications above. After modifying the skill, run scripts/sync-codex-skills.py to regenerate the Codex-compatible prompt at codex-prompts/deep-company-series.md.

Summary

• Select high-complexity tier in skills/deep-company-series.md (lines 26-30) to unlock the 8-article format for conglomerate analysis
• Implement all eight pillars from the template (lines 42-53), mapping each to a numbered markdown file
• Target specific word counts: 4k-5k for intro, 6k-8k for body articles, 10k-12k for valuation conclusion
• Create 00-系列说明.md as the internal index and store articles in dated folders under reports/{公司名}/
• Enforce structural standards: header blocks, data-first hooks, key-point reviews, and next-episode teasers
• Validate quality using the 7-item checklist and Explore agent cross-article verification
• Maintain style compliance with mobile-optimized openings and privacy grep checks before release

Frequently Asked Questions

What distinguishes the 8-article configuration from shorter series?

The 8-article configuration activates the "high-complexity" path designed for companies with multiple business lines, hidden assets, and rich management history. According to skills/deep-company-series.md (lines 26-30), shorter series (3-5 articles) use condensed pillar combinations, while the 8-article format dedicates individual articles to hidden assets, AI variables, and management history that simpler corporate structures do not require.

How does the Explore agent validate cross-article consistency?

The Explore agent automatically scans all eight markdown files in the series directory to verify numerical consistency, standardized terminology, and cross-reference integrity (lines 63-70). It ensures that financial figures cited in the profit engine article align with the valuation article, and that management quotes in article 7 do not contradict historical data in article 1.

Can I customize the word-count targets for specific articles?

While the skill recommends 4,000–5,000 characters for the introduction, 6,000–8,000 for middle pillars, and 10,000–12,000 for the final valuation piece (lines 46-53), these ranges serve as guidance rather than hard limits. However, deviating significantly from these targets may break the "textbook-level" pacing expected by readers of the 8-article format.

What happens if a previous series exists for the same company?

The "目录冲突规则" (directory conflict rules) in lines 58-60 prohibit overwriting existing series. When a folder matching reports/{公司名}/《看懂{公司名}》-{YYYYMMDD}/ already exists, you must generate a new timestamped directory rather than appending to or replacing the previous analysis. This preserves historical versions and prevents accidental data loss.

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