Industry-Funnel vs Industry-Research Skills: AI-Berkshire Filtering Methodology Explained

Industry-funnel narrows a full-market universe to 3-5 high-conviction stocks through quantitative screens and moat analysis, while industry-research constructs a complete industry-chain panorama to build diversified sector portfolios across all market tiers.

The ai-berkshire repository implements these distinct filtering methodologies as structured "skills" for systematic equity analysis. Both approaches guide investors through rigorous analytical frameworks, but they differ fundamentally in scope, execution steps, and deliverables. Understanding when to apply each—or how to combine them—determines whether you generate a concentrated watchlist or a comprehensive sector allocation strategy.

Primary Goals: Company Selection vs Industry Mapping

The essential distinction lies in the analytical end-state each skill produces.

Industry-funnel adopts a company-centric approach designed to produce a concise shortlist for deep investment analysis. It progressively eliminates candidates through hard filters until only 3-5 target stocks remain, each accompanied by investment theses, position sizing, and risk notes.

Industry-research adopts a chain-centric approach that evaluates the entire sector before any narrowing occurs. It constructs a complete industry-chain panorama covering upstream, midstream, and downstream segments across A-share, Hong Kong, US, and international markets. The output is a sector-level investment portfolio with tiered allocations (core, satellite, option, and ETF substitutes).

Filtering Mechanics: The Four-Layer Funnel vs Chain-Centric Scan

Each methodology follows a distinct logical sequence encoded in the repository's skill definitions.

Industry-Funnel: Multi-Stage Quantitative Screens

As defined in skills/industry-funnel.md#L13-L33, the funnel applies a four-layer elimination process:

1. Full-market scan – Screens active stocks by momentum and market-cap to generate an initial universe of 30-60 firms.
2. Hard-value criteria – Applies five quantitative filters (PE, ROE, cash-flow, debt-ratio, and moat/护城河 assessment) to reduce the list to 10 or fewer candidates.
3. Granular analysis – Requires 300-500 word qualitative write-ups to verify business quality, further pruning the list.
4. Four-masters deep dive – Applies intensive fundamental analysis to arrive at a final shortlist of 3-5 companies.

Each layer explicitly records why companies were eliminated, creating an audit trail for the final selection.

Industry-Research: Logic-Chain Construction

According to skills/industry-research.md#L16-L30, the research methodology follows a mapping workflow:

1. Logic-chain construction – Establishes trend → demand → bottleneck → sector relationships to identify structural drivers.
2. Full-chain scan – Captures all listed companies across multiple exchanges that participate in the industry value chain.
3. Tiered classification – Segments companies into Tier 1-4 categories based on market-cap, purity of exposure, and segment leadership.
4. Four-masters analysis – Applies deep fundamental assessment only to Tier 1-2 firms within each segment.

This approach ensures no relevant sub-segment is overlooked before portfolio construction begins.

Output Formats and Deliverables

The structural differences produce markedly different deliverables.

Industry-funnel generates a targeted investment report limited to 3-5 stocks. Each entry includes specific investment theses, recommended position sizing, and detailed risk notes. The report also maintains a record of elimination rationale at each funnel stage.

Industry-research generates a comprehensive industry-chain report containing segment tables, tiered company listings across all relevant exchanges, and overall portfolio allocation frameworks. This enables investors to understand the whole sector and allocate across multiple sub-segments rather than concentrating on individual securities.

Practical Implementation: Running the Skills

Both skills are accessible via CLI commands within the ai-berkshire framework. The repository processes these through scripts/run_skill.py and outputs markdown reports to the reports/ directory.

To execute an industry-funnel analysis for "AI算力" (AI computing power):

/industry-funnel AI算力

To execute a full industry-research scan for "创新药" (innovative drugs):

/industry-research 创新药

For programmatic access using the repository's internal CLI wrapper:

import subprocess

# Run industry funnel for robotics (机器人)

subprocess.run(["python3", "scripts/run_skill.py", "industry-funnel", "--arg", "机器人"])

# Run industry research for nuclear power (核电)

subprocess.run(["python3", "scripts/run_skill.py", "industry-research", "--arg", "核电"])

These commands generate timestamped markdown reports (e.g., reports/机器人-funnel-20260707.md) that include the full analytical workflow, data tables, and tier classifications. The tools/report_audit.py validation pipeline verifies all data before publication, ensuring consistency across both methodologies.

Complementary Workflow Integration

While each skill functions independently, the ai-berkshire architecture designs them as sequential complements rather than alternatives.

Industry-research is typically used first to define the competitive arena and identify all relevant sub-segments. Industry-funnel then converts this broad sector view into a concrete, high-conviction stock shortlist. Conversely, running the funnel first may reveal gaps in sub-segment coverage, prompting a return to the research skill to ensure the shortlist captures the full industry chain.

This complementary relationship ensures that analysis of "head companies" (龙头企业) in the funnel is grounded in comprehensive sector understanding derived from the research phase.

Summary

• Industry-funnel (skills/industry-funnel.md) implements a company-centric, multi-stage filter that reduces thousands of stocks to 3-5 high-conviction targets through quantitative screens and moat assessment.
• Industry-research (skills/industry-research.md) implements a chain-centric mapping system that evaluates all listed players across markets to construct complete industry-chain portfolios.
• Execution occurs via CLI commands or Python wrappers that generate timestamped reports in the reports/ directory, validated by tools/report_audit.py.
• Workflow optimization typically sequences research before funnel to ensure comprehensive coverage, though each skill operates independently.
• Deliverables differ fundamentally: the funnel produces concentrated investment theses, while research produces diversified sector allocation frameworks.

Frequently Asked Questions

Can I use industry-funnel without first running industry-research?

Yes, the funnel operates as a standalone screening tool that can process any full-market universe. However, without the prior industry-research phase, you risk overlooking relevant sub-segments or selecting companies from industry chains you haven't fully mapped. According to the ai-berkshire methodology, using research first ensures the subsequent funnel analysis covers all structural components of the sector.

What are the "four-masters" mentioned in both methodologies?

The "four-masters" (四大师) framework represents the deep qualitative analysis step applied in both skills, though at different stages. It refers to a comprehensive fundamental assessment of business quality, competitive advantages, management capability, and valuation metrics. In industry-funnel.md, it serves as the final filter before selection, while in industry-research.md, it validates Tier 1-2 companies within each industry segment.

How many companies typically remain after each funnel stage?

The skills/industry-funnel.md file specifies a typical progression: 30-60 companies after the initial full-market scan, fewer than 10 after applying the five hard-value criteria (PE, ROE, cash-flow, debt-ratio, and moat), and a final shortlist of 3-5 companies after the four-masters deep dive. Each stage documents elimination rationale to maintain analytical rigor.

Where are the generated reports stored in the ai-berkshire repository?

Both skills output markdown reports to the reports/ directory with standardized naming conventions. For example, running the funnel on "机器人" generates reports/机器人-funnel-20260707.md, while research outputs follow similar timestamped patterns. Before finalization, tools/report_audit.py validates data integrity and cross-references against the codex-prompts defined in codex-prompts/industry-funnel.md and codex-prompts/industry-research.md.

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