Handling Information Scarcity in Private-Company Research: A Multi-Agent AI Framework

The ai-berkshire repository solves information scarcity in private-company research through a multi-agent workflow that cross-validates data across seven specialized roles, applies confidence scoring (🟢🟡🔴), and falls back to first-principles analysis when data is extremely limited.

Analyzing private companies presents unique challenges due to the absence of public financial disclosures and standardized reporting. The xbtlin/ai-berkshire repository addresses handling information scarcity in private-company-research through a systematic, agent-based methodology defined in skills/private-company-research.md, ensuring transparent, bias-aware analysis even when traditional data sources are unavailable.

The Core Challenge of Information Scarcity

Private companies lack SEC filings and audited financial statements, forcing analysts to stitch together fragmented data from disparate sources. The workflow operates on the core philosophy that “absence of data ≠ poor business” and mandates strict transparency about verified facts versus inferred reasoning. Rather than fabricating detail, the system classifies information zones as 白区 (known), 灰区 (partial), or 黑区 (unknown) to highlight critical blind spots.

Multi-Agent Architecture for Private Company Analysis

The system organizes research across seven specialized agents orchestrated by a central team-lead. According to the execution flow defined in skills/private-company-research.md, the team-lead initiates the workflow via a TeamCreate call, then issues six TaskCreate commands (one per specialized role) containing detailed subjects, active forms, and deliverable specifications.

The Seven Specialist Roles

Each agent focuses on a distinct research dimension to maximize coverage of scarce information:

• team-lead: Orchestrates workflow, validates assembled data, and produces the final investment decision report.
• business-decoder: Dissects the business model, product matrix, and user ecosystem to answer “What does the business actually do?”
• financial-detective: Reconstructs financials, estimates valuation, and tracks cash-burn to assess “How does the business look financially when data is missing?”
• competitive-mapper: Maps industry structure and substitution threats to identify rivals and disruption risks.
• risk-governance-analyst: Evaluates governance quality, team capability, and legal/regulatory exposure.
• tech-ip-analyst: Analyzes technology stack, patents, and R&D capability to identify durable moats.
• signal-miner: Mines alternative data including recruiting trends, app metrics, and social sentiment for hidden signals beyond traditional news.

The workflow runs these agents in parallel as required by the “第七步：启动6个并行Agent” clause, then aggregates results through a structured reconciliation process.

Hierarchical Data Sourcing and Confidence Scoring

To combat information scarcity systematically, the skill implements a priority-ordered source hierarchy that prioritizes high-credibility documents over speculative intelligence.

The Credibility Hierarchy

The prescribed data-sourcing matrix follows this descending order of reliability:

1. SEC filings (招股书) — official regulatory documents
2. Parent-company reports (母公司年报) — consolidated financial disclosures
3. Regulatory disclosures (监管披露) — government filings
4. Bond documents (债券文件) — debt issuance disclosures
5. Business registrations (工商登记) — corporate registry data (e.g., 天眼查)
6. Funding news (融资新闻) — verified capital raise announcements
7. Third-party research (第三方研报) — analyst reports
8. In-depth media (深度媒体) — investigative journalism
9. Industry estimates (行业推算) — sector-based calculations
10. Former employee insights (前员工爆料) — crowdsourced intelligence

For each data point, analysts must record the source type, timestamp (year-month), and confidence level: 🟢 (high) for official filings, 🟡 (medium) for reputable media, and 🔴 (low) for crowdsourced rumors or estimates.

Cross-Validation Requirements

Every metric used in valuation or risk assessments must be cross-validated with at least two independent sources before inclusion. The “第六步：交叉验证与信息拼图” section mandates Data Conflict Arbitration, where the team-lead aligns overlapping metrics (e.g., revenue from SEC filings versus news reports) and documents the rationale for source selection.

First-Principles Mode for Extreme Scarcity

When the pipeline detects 极度稀缺 (extreme scarcity)—defined as fewer than three credible data sources for a core metric—the workflow automatically switches to a first-principles mode. Rather than forcing incomplete data into standard templates, this fallback answers four essential strategic questions:

1. Problem validation: Is the demand real and persistent?
2. Team advantage: Why can this specific team capture the market?
3. Upside ceiling: What is the potential market size if successful?
4. Validation milestones: What specific evidence would prove or disprove the investment thesis?

This mode is documented under “应对原则 → 第一性原理模式” in skills/private-company-research.md and prevents the analytical errors that arise from fabricated precision.

Cross-Agent Validation and Signal Consistency

After the six agents submit sub-reports, the team-lead performs a three-step reconciliation to ensure narrative coherence:

1. Data Conflict Arbitration: Resolves discrepancies between sources (e.g., revenue figures from 招股书 versus industry reports) with documented rationale.
2. Signal Consistency Check: Validates that the business-decoder’s growth narrative aligns with the signal-miner’s recruiting trend data, and that the tech-ip-analyst’s patent moat assessment matches the competitive-mapper’s market positioning.
3. Information-Gap Mapping: Explicitly catalogs remaining unknowns to prevent overconfidence in the final valuation.

Practical Implementation Guide

Invoking via Claude

To trigger the workflow in Claude or Claude-Code, use the slash command:

/private-company-research

The system automatically:

• Creates the team via TeamCreate
• Spawns six parallel tasks with embedded prompts from skills/private-company-research.md
• Aggregates sub-reports and writes the final output to reports/{公司名}/{公司名}-private-{YYYYMMDD}.md

Command-Line Installation

For local CLI usage, install the slash-commands via the provided script:

./scripts/install-claude-commands.sh
claude /private-company-research "蚂蚁集团"

This generates a team named ant-group-private-research and produces a report at reports/ant-group/ant-group-private-YYYYMMDD.md following the naming convention specified in the skill definition.

Programmatic Access via Codex

The repository provides a Codex-compatible prompt layer at codex-prompts/private-company-research.md for integration with automated coding environments:

import json, requests

payload = {
    "prompt": "/private-company-research",
    "variables": {"ARGUMENTS": "Stripe"}
}
response = requests.post("https://api.cohere.ai/v1/generate", json=payload)
print(response.json()["text"])

This mirrors the multi-agent workflow for API-based consumption, returning structured markdown reports with confidence annotations.

Summary

• Multi-agent specialization: Seven distinct roles (business-decoder, financial-detective, etc.) parallelize research to maximize information extraction from scarce sources.
• Hierarchical sourcing: A ten-tier credibility matrix (SEC filings → former employee insights) ensures data quality through source triage.
• Confidence transparency: Mandatory 🟢🟡🔴 labeling and cross-validation requirements prevent over-reliance on single-source claims.
• First-principles fallback: Automatic switching to strategic question mode when data is extremely limited avoids fabricated precision.
• Implementation flexibility: Available via Claude slash commands, CLI tools in scripts/install-claude-commands.sh, or programmatic Codex prompts.

Frequently Asked Questions

How does the system handle conflicting data from different sources?

The team-lead performs Data Conflict Arbitration during the “第六步：交叉验证与信息拼图” phase, aligning overlapping metrics (e.g., revenue from 招股书 versus news reports) and documenting the specific rationale for trusting one source over another. This ensures the final valuation rests on explicitly justified data choices rather than hidden assumptions.

What happens when no reliable financial data exists for a private company?

When the pipeline detects fewer than three credible sources for core metrics (极度稀缺), it switches to first-principles mode as defined in “应对原则 → 第一性原理模式”. Instead of fabricating financial details, the analyst answers four strategic questions about problem validation, team advantage, market ceiling, and validation milestones, ensuring useful output despite minimal inputs.

How is the confidence level determined for each data point?

Confidence derives strictly from source type: 🟢 (high) for official filings and direct disclosures, 🟡 (medium) for mainstream media and reputable industry reports, and 🔴 (low) for calculations, rumors, or non-structured signals. Inferred values are italicized and accompanied by the specific 推算方法 (estimation methodology), such as “基于行业平均 LTV×用户数推算”.

Can this workflow be integrated with automated AI coding assistants?

Yes. The repository provides dual interfaces: Claude slash commands via scripts/install-claude-commands.sh for interactive use, and a Codex prompt layer at codex-prompts/private-company-research.md for API-based integration. Both implementations execute the same multi-agent workflow with confidence scoring and cross-validation.

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