How to Implement Private Company Research for Unlisted Companies in AI Berkshire

AI Berkshire implements private company research through a multi-agent workflow defined in skills/private-company-research.md, where a team-lead agent orchestrates six specialist agents to gather, validate, and synthesize data into a comprehensive markdown report.

AI Berkshire provides a systematic approach to analyzing unlisted firms through its dedicated research workflow. The repository's canonical skill file defines a parallel deep-dive methodology that transforms fragmented public signals into coherent investment theses. This implementation leverages Claude-Code and Codex command generators to automate the creation of research teams and task distribution.

Understanding the Multi-Agent Architecture

The private company research workflow operates as a three-layer system designed specifically for unlisted entities that lack public filings.

Team Orchestration Layer

At the core of the implementation is the TeamCreate mechanism (lines 55-60 in skills/private-company-research.md). The team-lead agent instantiates a dedicated research team named {company}-private-research and spawns six background agents running in parallel. Each agent operates as a general-purpose type (lines 92-94) with specialized instructions tailored to unlisted company analysis.

The Six Specialist Agents

The workflow distributes cognitive labor across distinct functional domains:

• business-decoder – Analyzes business models and user metrics
• financial-detective – Stitches financial data and performs valuation calculations
• competitive-mapper – Maps industry landscapes and competitive positioning
• risk-governance-analyst – Evaluates management quality and governance structures
• tech-ip-analyst – Assesses technology stacks and intellectual property
• signal-miner – Extracts alternative data from recruiting platforms, patent filings, and app store metrics (lines 53-54, 61-68)

Configuring Task Definitions and Validation Rules

Each specialist receives a TaskCreate payload containing a subject, activeForm instruction, and detailed description specifying required data points and table formats.

Parallel Task Execution

The system launches six simultaneous TaskCreate operations (lines 61-68) to maximize research speed. Each task block includes enumerated requirements (Task 1 through Task 6) that drive agents to gather specific datasets in parallel, reducing total research time for unlisted companies.

Confidence Rating and Cross-Validation

Every data point requires explicit confidence tagging using a three-tier system defined in lines 27-28 and 26-29:

• 🟢 High confidence: Verified by dual sources
• 🟡 Medium confidence: Single reliable source or reasonable inference
• 🔴 Low confidence: Speculative or unverified data

The team-lead performs mandatory cross-validation (lines 31-34, 52-55) to arbitrate conflicts between agent reports before final synthesis.

Invoking the Private Company Research Workflow

Users can trigger the workflow through multiple interfaces provided by the AI Berkshire runtime.

Slash Command Interface

The simplest invocation uses the Claude or Codex client:

/private-company-research SpaceX

This expands automatically into the full workflow, creating the space-x-private-research team and returning progress updates through the chat interface.

Direct API Integration

For programmatic access, POST to the internal skill endpoint:

import json
import requests

payload = {
    "skill": "private-company-research",
    "arguments": "Ant Group"
}

resp = requests.post(
    "http://localhost:8000/api/skill",
    json=payload
)

print(json.dumps(resp.json(), indent=2))

The server triggers team creation, dispatches parallel agents, and writes the final report to reports/ant-group/ant-group-private-20260711.md.

Synchronizing Skill Changes

After modifying skills/private-company-research.md, regenerate the Codex artifacts:

python3 scripts/sync-codex-skills.py

This updates codex-skills/private-company-research/SKILL.md and codex-prompts/private-company-research.md with the latest task definitions.

Report Structure and Output Format

The final synthesis occurs in the "information-puzzle" step (lines 52-66, 70-78), where the team-lead merges six sub-reports into a unified markdown document.

The output typically follows this structure:

#### 1. 一句话结论

> Ant Group 的真实价值约在 $180-$220B 之间，核心驱动是其支付网络的规模效应与数据壁垒。

| 项目 | 内容 | 置信度 |
|---|---|---|
| 公司名称 | Ant Group | 🟢 |
| 成立时间 | 2004年 | 🟢 |
| 最新估值 | $200B（2023轮） | 🟢 |
| 关键用户规模 | 1.2B MAU | 🟡 |
| 主要投资方 | 高瓴、软银等 | 🟢 |

Reports are saved to reports/{Company}/{Company}-private-{YYYYMMDD}.md according to the "保存报告" section of the skill file.

Summary

• AI Berkshire implements private company research through a multi-agent workflow defined in skills/private-company-research.md
• The system orchestrates six specialist agents (business-decoder, financial-detective, competitive-mapper, risk-governance-analyst, tech-ip-analyst, and signal-miner) working in parallel
• Confidence tagging (🟢/🟡/🔴) and cross-validation ensure data quality for unlisted entities lacking public filings
• Use /private-company-research {Company} for CLI invocation or POST to /api/skill for programmatic access
• Run python3 scripts/sync-codex-skills.py after editing skill definitions to update generated artifacts

Frequently Asked Questions

What file contains the canonical definition of the private company research workflow?

The primary implementation resides in skills/private-company-research.md at the repository root. This file contains the TeamCreate logic, task definitions for all six specialist agents, confidence rating conventions, and cross-validation rules. After editing this file, you must run scripts/sync-codex-skills.py to regenerate the Codex-compatible versions in codex-skills/private-company-research/SKILL.md.

How does AI Berkshire handle data quality for unlisted companies without public financials?

The workflow enforces dual-source verification through the team-lead's cross-validation step (lines 31-34). Every key metric receives a confidence rating (high/medium/low), and the signal-miner agent specifically targets alternative data sources like recruiting activity, patent filings, and app store metrics to triangulate financial estimates when traditional filings are unavailable.

Can I customize the six specialist agents for specific industries?

Yes. The activeForm parameters within each TaskCreate payload in skills/private-company-research.md define the instructions for each agent. You can modify these task blocks to add industry-specific data requirements or adjust the validation criteria. Remember to run python3 scripts/sync-codex-skills.py after making changes to ensure the Codex runtime receives the updated prompts.

Where are the final research reports stored?

The system saves completed analyses to reports/{Company}/{Company}-private-{YYYYMMDD}.md as specified in the skill's "保存报告" section. This path structure organizes outputs by company name and research date, making it easy to track historical valuations for unlisted companies over time.

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