# How to Use the /private-company-research Skill for Unlisted Companies in AI Berkshire

> Discover how to use the private-company-research skill to generate deep research reports for unlisted companies like SpaceX. This AI Berkshire tool deploys specialist agents to gather and cross-validate scarce public data.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: how-to-guide
- Published: 2026-07-25

---

**The `/private-company-research` skill orchestrates a parallel multi-agent workflow that generates deep-research reports for unlisted companies like SpaceX or Ant Group by deploying six specialist agents to gather and cross-validate scarce public data.**

The `xbtlin/ai-berkshire` repository provides a sophisticated research framework for financial analysis, with the `/private-company-research` skill serving as the primary entry point for investigating non-public entities. This skill implements a three-layer architecture that automatically manages team creation, parallel task execution, and conflict arbitration to produce structured markdown reports. Because private companies lack SEC filings and traditional disclosure requirements, the skill deploys specialized agents to extract intelligence from web sources, industry reports, and competitive landscapes.

## What is the /private-company-research Skill?

The `/private-company-research` skill is a canonical deep-research command defined in **[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)**. It functions as the source of truth for investigating companies where public data is limited or fragmented, such as pre-IPO startups and privately-held enterprises. Unlike standard research tools, this skill creates a **team-lead** agent that coordinates six domain specialists simultaneously, ensuring comprehensive coverage of business model, financial health, competitive positioning, risk factors, intellectual property, and market signals.

The skill is referenced in the Skills Overview table of [**[`README_EN.md`](https://github.com/xbtlin/ai-berkshire/blob/main/README_EN.md)** (lines 69-75)](https://github.com/xbtlin/ai-berkshire/blob/main/README_EN.md), which documents it as one of 20 available entry points in the AI Berkshire ecosystem. Any modifications to the research workflow must be made directly in the canonical skill file and propagated using the [`scripts/sync-codex-skills.py`](https://github.com/xbtlin/ai-berkshire/blob/main/scripts/sync-codex-skills.py) synchronization script, ensuring consistency across both Claude Code and Codex CLI interfaces.

## Three-Layer Architecture

The skill operates within the repository's established **Skill → Agent → Tool** hierarchy, as detailed in [**[`AGENTS.md`](https://github.com/xbtlin/ai-berkshire/blob/main/AGENTS.md)**](https://github.com/xbtlin/ai-berkshire/blob/main/AGENTS.md):

- **Skill Layer**: The command interface defined in **[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)** handles user input validation and workflow initialization. This layer determines which specialist agents to spawn and how to structure the final report.

- **Agent Layer**: Upon invocation, the system instantiates a team-lead agent that creates six specialist agents with specific `activeForm` configurations. These agents execute in parallel, each focusing on distinct research vectors such as business decoding, financial detective work, and competitive mapping.

- **Tool Layer**: Each specialist accesses lower-level tools including **web search**, **web fetch**, **report audit**, and source validation utilities. Agents automatically tag findings with confidence indicators (🟢 high, 🟡 medium, 🔴 low) during data collection.

The Codex wrapper located at **[`codex-prompts/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-prompts/private-company-research.md)** simply loads the skill definition from the generated **[`codex-skills/private-company-research/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/private-company-research/SKILL.md)** file, ensuring the same execution logic applies regardless of interface.

## How to Run the Skill

### Command Line Interface

You can invoke the research workflow through either Claude Code or the Codex CLI:

```bash

# Via Codex CLI (macOS / Linux)

codex --prompt "/private-company-research SpaceX"

# Via Claude Code slash command

/private-company-research Ant Group

```

Both methods trigger the identical multi-agent workflow and generate a timestamped report at `reports/<Company>/<Company>-private-<YYYYMMDD>.md`.

### Programmatic Invocation

For automated pipelines or Python scripts, execute the skill via subprocess:

```python
import subprocess
from datetime import datetime

def research_unlisted_company(company_name: str):
    """Trigger private company research via Codex CLI."""
    cmd = ["codex", "--prompt", f"/private-company-research {company_name}"]
    subprocess.run(cmd, check=True)
    print(f"Report generated for {company_name}")

# Example usage

research_unlisted_company("Stripe")

```

## The Six-Phase Research Workflow

According to the workflow specification in [**[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)** (lines 41-88)](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md), the skill executes through nine structured steps:

### 1. Team Initialization

The team-lead agent creates a dedicated research team named `<company>-private-research` with `agent_type: team-lead`, establishing the coordination framework for parallel execution.

### 2. Task Distribution

The system generates six concurrent tasks, each assigned to a specialist agent:
- **business-decoder**: Analyzes business models and product-user relationships
- **financial-detective**: Investigates revenue streams, funding history, and unit economics
- **competitive-mapper**: Maps market positioning and competitive dynamics
- **risk-governance-analyst**: Evaluates regulatory, legal, and governance risks
- **tech-ip-analyst**: Assesses technology stack and intellectual property portfolio
- **signal-miner**: Extracts market signals from news, hiring patterns, and industry chatter

### 3. Parallel Execution

All six agents execute simultaneously, using web search and fetch tools to gather data. Each agent operates with a specific `activeForm` guiding its research methodology, ensuring comprehensive coverage without duplication.

### 4. Cross-Validation

Agents verify findings against multiple sources, assigning confidence indicators to each data point. The system flags contradictions between agents for later resolution.

### 5. Conflict Arbitration

The team-lead reviews conflicting data points from parallel streams, determining authoritative values through source reliability weighting and timestamp analysis.

### 6. Report Synthesis

The team-lead aggregates all research into a final markdown document following the template in **[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)**. The report includes sections for business analysis, financials, competition, risks, technology, and market signals.

### 7. Resource Cleanup

Upon completion, the system executes the cleanup phase defined in **"第九步：清理团队"** (Step 9), deleting the temporary research team to free computational resources.

## Understanding the Output

The final deliverable is a comprehensive markdown report saved to `reports/<Company>/<Company>-private-<YYYYMMDD>.md`. This document contains:

- **Executive Summary**: Consolidated findings from all six research vectors
- **Confidence Metrics**: Color-coded reliability indicators (🟢🟡🔴) for each data point
- **Source Attribution**: URLs and references for all external data
- **Structured Analysis**: Dedicated sections matching the six agent specializations

After generation, the report remains available in the repository's `reports/` directory for downstream analysis or integration with other AI Berkshire skills.

## Summary

- The **`/private-company-research`** skill in `xbtlin/ai-berkshire` provides deep-research capabilities specifically designed for unlisted companies lacking traditional disclosure data.
- Execution requires only a single command (`/private-company-research <Company>`) via Claude Code or `codex --prompt` via CLI, triggering automatic parallel agent orchestration.
- The canonical workflow definition resides in **[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)**, which coordinates six specialist agents (business-decoder, financial-detective, competitive-mapper, risk-governance-analyst, tech-ip-analyst, signal-miner) through a team-lead architecture.
- Research outputs include confidence-tagged data and structured markdown reports saved to the `reports/` directory with standardized naming conventions.

## Frequently Asked Questions

### How does the skill handle companies with extremely limited public data?

The **`/private-company-research`** skill employs a **signal-miner** agent specifically designed to extract intelligence from indirect sources such as job postings, patent filings, supply chain announcements, and competitor disclosures. When primary data is unavailable, the workflow instructs agents to triangulate financial estimates from secondary sources and explicitly flag low-confidence findings with 🔴 indicators, ensuring transparency about data limitations while maximizing available intelligence.

### Can I modify the research workflow for specific industries?

Yes, but modifications must be made to the canonical source file **[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)** rather than the generated wrappers. After editing the skill definition—such as adjusting the `activeForm` prompts for specific industry verticals—you must run `python scripts/sync-codex-skills.py` to propagate changes to **[`codex-skills/private-company-research/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/private-company-research/SKILL.md)**. This synchronization ensures both Claude Code and Codex CLI users receive identical behavior.

### What is the difference between using Claude Code versus Codex CLI for this skill?

Both interfaces invoke the identical underlying workflow defined in the canonical skill file. Claude Code users access the skill via the slash command `/private-company-research`, while Codex CLI users execute `codex --prompt "/private-company-research <Company>"`. The Codex implementation uses a minimal wrapper at **[`codex-prompts/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-prompts/private-company-research.md)** that loads the full skill definition, ensuring functional parity across platforms.

### How long does the research process typically take for a single unlisted company?

Execution time varies based on data availability and web search latency, but the parallel architecture typically completes the full six-agent workflow in 5-10 minutes for companies with moderate public footprints (e.g., established unicorns like SpaceX). Companies with minimal digital presence may complete faster due to fewer data sources, while complex conglomerates like Ant Group may require additional time for the risk-governance-analyst to parse regulatory complexity. The team-lead agent monitors completion status and proceeds to synthesis only after all six specialists report completion or timeout.