How the /industry-funnel Skill Filters Stocks: A 4-Layer Value Investing Architecture

The /industry-funnel skill implements a systematic four-stage quantitative funnel that narrows a broad market universe down to three high-conviction stocks through layered screening, hard-filter criteria, structured granular analysis, and master-style deep dives.

The /industry-funnel skill in the xbtlin/ai-berkshire repository provides a rigorous, reproducible framework for value-oriented stock selection based on Berkshire Hathaway investment principles. This skill applies a multi-layer filtering architecture that combines quantitative metrics with qualitative deep analysis to identify exceptional opportunities across A-share, HK, US, and international markets.

The Four-Layer Funnel Architecture

According to the source code in skills/industry-funnel.md, the skill processes stocks through four sequential layers, each with explicit inclusion criteria and mandatory audit trails. This design ensures that every elimination decision is documented and reproducible.

Layer 1: Universe Screening (全市场扫描)

The first layer collects all active tickers from A-share, HK, US, and international markets, including future IPO candidates. As implemented in lines L41-L55 of skills/industry-funnel.md, this stage merges three distinct sub-sets:

• A-type: Stocks ranked by 30-day average turnover
• B-type: Top-20 gainers over 30-day and 90-day periods
• C-type: Sector-wide top-30 companies by market capitalization

The union of these categories creates an initial pool of 30-60 tickers that advance to quantitative filtering.

Layer 2: Value Investing Hard Filters (价值投资硬指标粗筛)

The second layer applies five quantitative "hard" filters to every candidate from Layer 1. Located at lines L82-L90 in skills/industry-funnel.md, this stage requires all five criteria to pass, or alternatively four passes plus one "near-pass" (which must be explicitly highlighted).

If more than 12 companies survive this filter, the moat threshold tightens from ★★★★ to ★★★★ and the filter repeats until the pool is sufficiently narrow. This dynamic tightening mechanism ensures the funnel maintains a manageable cohort for deeper analysis.

Layer 3: Granular Analysis (精细分析)

For the 10 or fewer survivors of Layer 2, the skill generates a structured 300-500-word analysis following a strict template defined in lines L13-L45. Each analysis covers:

• Business model assessment
• Financial quality evaluation
• Moat depth examination using the five-category taxonomy (brand, switching cost, network effect, scale, technology/license)
• Top-3 risk identification
• Valuation quick-check

Layer 4: Four-Master Deep Dive (四大师深度分析)

The final stage subjects the top candidates to an 800-1200-word "four-master" analysis (段永平, 巴菲特, 芒格, 李录) as specified in lines L46-L53. Rather than purely rank-based selection, this layer enforces portfolio complementarity, ensuring the final three picks represent distinct risk profiles: core-stable, growth-mid, and high-beta.

Governance, Audit Controls, and Bias Mitigation

The /industry-funnel skill incorporates explicit governance mechanisms to prevent black-box decisions. Every company removed at any stage must have a documented "淘汰理由" (elimination reason), creating a complete audit trail.

The workflow includes dedicated checks against:

• AI-story bias: Preventing narrative fallacies in automated research
• English-source bias: Ensuring non-English market coverage receives equal weight
• Market-type bias: Balancing large-cap incumbents against IPO candidates

The five-category moat taxonomy uses a ★-rating system throughout the pipeline, providing consistent evaluation criteria across all layers.

How to Invoke the /industry-funnel Skill

You can execute the skill through CLI commands or programmatic Python wrappers.

Run the skill for a specific sector:

/industry-funnel AI Compute

Programmatic invocation using the Python wrapper:

from ai_berkshire import invoke_skill

result = invoke_skill(
    "industry-funnel",
    args={"行业": "AI Compute"},
    output_path="reports/AI-Compute-funnel-20260729.md"
)
print(result['summary'])

Debug a specific layer using the audit tool:

# Layer 2 – hard-filter check

python3 tools/report_audit.py extract \
    --report reports/AI-Compute-funnel-20260729.md \
    --stage hard_filter

Output Contract and Deliverables

The final report is written to reports/{行业名}-funnel-{YYYYMMDD}.md and must include data source links, estimation flags, and a self-assessment matrix. The tools/report_audit.py script validates the final report against mandatory data-source checklists to ensure compliance with the skill's quality standards.

Summary

• The /industry-funnel skill uses a four-layer sequential architecture defined in skills/industry-funnel.md to filter thousands of stocks down to three high-conviction picks
• Layer 1 creates a 30-60 ticker universe from A-share, HK, US, and IPO candidates using turnover, momentum, and market-cap criteria
• Layer 2 applies five value-investing hard filters with dynamic threshold tightening when results exceed 12 survivors
• Layer 3 generates 300-500 word structured analyses covering moats, risks, and valuations for up to 10 finalists
• Layer 4 conducts 800-1200 word "four-master" deep dives, enforcing portfolio complementarity across the final three selections
• Mandatory audit trails require documented elimination reasons for every rejected stock, ensuring reproducibility and bias mitigation

Frequently Asked Questions

What makes the /industry-funnel skill different from standard stock screeners?

Unlike static screeners, the /industry-funnel skill implements dynamic threshold tightening and mandatory narrative analysis. When Layer 2 produces more than 12 survivors, the moat threshold automatically tightens from ★★★★ to ★★★★ and re-runs. Furthermore, every elimination requires a documented reason ("淘汰理由"), creating an audit trail impossible in traditional black-box filters.

How does the skill prevent AI bias in stock selection?

The skill includes explicit bias checks against AI-story bias, English-source bias, and market-type bias (large-cap versus IPO candidates). These safeguards are documented in the "AI 研究偏见自觉" section of skills/industry-funnel.md, ensuring the analysis weights non-English sources and smaller market candidates appropriately rather than favoring familiar large-cap narratives.

What are the five value investing hard filters in Layer 2?

While the specific quantitative thresholds are configurable, the five hard filter categories evaluate: (1) Profitability metrics, (2) Financial stability ratios, (3) Capital efficiency, (4) Earnings quality, and (5) Valuation bounds. Companies must pass all five criteria or four plus one explicitly flagged near-pass to advance. If more than 12 companies qualify, the moat rating threshold tightens and the layer re-executes.

Where are the final reports stored and how are they validated?

Final reports are written to reports/{行业名}-funnel-{YYYYMMDD}.md following a strict output contract that includes data source links and estimation flags. The tools/report_audit.py utility validates these reports against mandatory checklists, verifying that every claim is sourced and every elimination has a documented reason, ensuring the deliverable meets the skill's reproducibility standards.

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