# How to Use the Industry-Funnel Skill in AI-Berkshire for Value Investing Research

> Master the industry-funnel skill in AI-Berkshire to refine your value investing research. Discover how this quantitative tool narrows themes to three top candidates using Claude Code or Codex.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: how-to-guide
- Published: 2026-07-10

---

**The industry-funnel skill narrows a broad thematic search to three final investment candidates through a four-stage quantitative filter, invoked via `/industry-funnel <theme>` in Claude Code or the Codex prompt wrapper.**

The **industry-funnel skill** is a declarative workflow defined in the `xbtlin/ai-berkshire` repository that transforms broad industry themes into rigorously vetted investment shortlists. Unlike traditional screening tools, this skill implements a multi-stage funnel process that combines quantitative hard filters with qualitative deep-dive analysis, outputting auditable markdown reports ready for value-oriented decision making.

## Four-Stage Quantitative Workflow

The skill orchestrates a disciplined progression from market scan to final selection, documented entirely in [`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md).

### Stage 1: Market Scan

Collect all relevant tickers across A-shares, Hong Kong, US exchanges, and pre-IPO candidates using data sources like 同花顺, 东方财富, 富途, and NASDAQ.

### Stage 2: Hard-Filter

Apply five quantitative criteria to eliminate weak candidates:

- **PE ratio** thresholds
- **ROE** minimums
- **Cash-flow** stability
- **Debt-ratio** caps
- **Moat rating** of ★★★ or higher

### Stage 3: Structured Analysis

Generate a 300-500 word structured review for each surviving company, analyzing business model durability and competitive positioning.

### Stage 4: Deep-Dive Selection

Produce an 800-1200 word "Four Masters" analysis for the final three picks, evaluating investment portfolio complementarity and margin of safety.

## Invocation Methods

The skill exposes identical functionality through two interfaces that both resolve to the same underlying markdown workflow.

### Claude Code Slash Command

Type the command directly in the Claude UI:

```text
/industry-funnel AI 算力

```

The parser substitutes `$ARGUMENTS` with your input and executes the four-stage workflow defined in [`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md).

### Codex Prompt Wrapper

For Codex users, invoke via the generated prompt:

```bash
python3 -m codex.run codex-prompts/industry-funnel.md "AI 算力"

```

This loads [`codex-prompts/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-prompts/industry-funnel.md), which forwards the request to the master skill file.

## Auditing Generated Reports

After the funnel produces a report in `reports/{行业}-funnel-{YYYYMMDD}.md`, verify data accuracy using the audit tool.

Extract data points for verification:

```bash
python3 tools/report_audit.py extract --report reports/AI算力-funnel-20260710.md

```

Fill the generated JSON template with fetched values from macrotrends, eastmoney, and other sources, then run:

```bash
python3 tools/report_audit.py verdict --results '<filled-json>'

```

A **【准出】** verdict indicates the report cleared validation and is ready for publication.

## Key Architecture and Source Files

Understanding the file structure ensures correct usage and debugging:

- **[`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md)** – Master workflow definition containing stage descriptions, filter thresholds, and output table schemas.
- **[`codex-prompts/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-prompts/industry-funnel.md)** – Generated prompt wrapper that exposes the skill to Codex runners.
- **[`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py)** – Extraction and verification engine that cross-checks financial figures against external data providers.
- **`reports/`** – Dynamic output directory storing completed funnel analyses with standardized naming conventions.

## Summary

- The **industry-funnel skill** implements a four-stage filter (Market Scan → Hard-Filter → Structured Analysis → Deep-Dive) defined in [`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md).
- Invoke via **`/industry-funnel <theme>`** in Claude Code or the Codex prompt wrapper for identical results.
- The skill contains no executable code; it provides a declarative workflow for human or AI execution.
- Validate outputs using **[`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py)** to ensure financial data accuracy before publishing.
- Final reports follow strict templates in `reports/` with naming convention `{行业}-funnel-{YYYYMMDD}.md`.

## Frequently Asked Questions

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

Unlike automated screeners that only handle quantitative data, the industry-funnel skill mandates a 300-500 word structured qualitative analysis for surviving candidates and an 800-1200 word "Four Masters" deep-dive for final selections, ensuring both numerical and business-quality validation.

### Can I modify the filter thresholds in the hard-filter stage?

Yes. The thresholds for PE, ROE, cash-flow, debt-ratio, and moat ratings are defined in [`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md). Edit this markdown file to adjust quantitative criteria before running the funnel.

### How do I know if the generated report data is accurate?

Run `python3 tools/report_audit.py extract` to pull all financial claims from the report into a verification JSON template. After cross-referencing with external sources like macrotrends and eastmoney, submit the filled JSON to `report_audit.py verdict` to receive a **【准出】** clearance or flagged discrepancies.

### Does the skill automatically execute trades or fetch real-time data?

No. The industry-funnel skill is purely declarative and contains no executable trading logic. It structures a workflow for data gathering and analysis, but requires manual execution or AI assistance to populate the research stages and verify data through the audit tool.