Industry Funnel (AI Berkshire)
From whole market to 3 picks: a value-investing funnel with 5 hard criteria and four-master deep analysis.
The skill explicitly requires data to be sourced, estimates labeled, and includes a data sampling process for quality control. The repository contains a SECURITY.md providing a private channel for reporting vulnerabilities, indicating security awareness. However, the skill relies on external web searches and does not mandate user confirmation or least-privilege access controls for tools. Deductions for lacking clear permission minimization and user confirmation mechanisms.
The skill flow is self-consistent, from market-wide scanning to final selection of 3 companies. The repo includes tests (test_financial_rigor.py, test_report_audit.py) covering key paths of the tools, showing fixes for issues like GBK encoding and negative number extraction. But static review cannot execute these, and tests don't cover the full skill workflow, only the tools. Deductions for no execution and limited test coverage.
The skill's application scenario is clear, focused on industry funnel screening, and explicitly distinguishes from other skills like industry-research. Documentation supports Chinese, and the repo is oriented to Chinese users. However, core functionality depends on external data sources (e.g., 同花顺, 富途, Yahoo Finance) which may be inaccessible from mainland China, impacting real-world usability. Deductions for not addressing network restrictions alternatives.
Skill documentation is well-structured: name, description, applicable scenarios, funnel overview, step-by-step guidance, output templates, and quality gate. Repo has MIT license, detailed README, clear installation instructions, and test files. But lacks a changelog and version number, with maintenance responsibility on a single individual, and unclear update path. Deductions for missing versioning and systematic maintenance/update documentation.
Skill's goal is clear, with detailed process and output formats, directly producing reports for users. But static review cannot verify actual effectiveness, and report quality depends on external data availability and accuracy. The README includes practical examples (e.g., AI industry screening on 2026-05-09) providing some evidence, but actual reports aren't available to verify. Deductions for no execution and uncertainty due to external dependencies.
Repo includes test code as partial evidence of reproducibility, and README shows practical examples. But static review cannot independently verify report content correctness, and tests cover only tools, not the full skill. Deductions for static review limitations and incomplete verification coverage.
- Core functionality depends on external data sources (e.g., 同花顺, 富途, Yahoo Finance) which may be inaccessible from mainland China; confirm alternative solutions.
- Skill lacks explicit user confirmation and least-privilege controls; review commands and network requests before use to avoid overreach.
- Missing changelog and versioning; maintenance and update path unclear, reliant on individual maintainer.
- Static review cannot verify actual effectiveness; refer to README's practical examples or try it yourself.
What does this skill do, and when should you use it?
This is a value-investing screening framework based on the methodologies of Buffett, Munger, Duan Yongping, and Li Lu. It narrows research scope through a funnel: scan the entire market to 30-60 companies, apply five hard criteria to cut down to 10 or fewer, then analyze each in detail, and finally select 3 for deep four-master perspective analysis. It emphasizes keeping elimination reasons at every layer and selects finalists based on portfolio complementarity rather than pure scoring, mitigating AI biases like large-cap, English-language, and story-driven preferences. The skill works with both Claude Code and Codex, integrated into the AI Berkshire repo, sharing financial rigor tools (e.g., market-cap verification, multi-source cross-validation).
Executes an industry funnel value-investing screening. Given an industry or direction (e.g., 'AI compute'), it: first scans the whole market (A-shares, HK, US, plus unlisted candidates) using union of activity, gainers, and market-cap top-30 lists to get 30-60 names; then screens each with five hard criteria (PE valuation, ROE, operating cash flow, debt ratio, moat rating) down to <=10; performs 300-500-word structured analysis per company (business model, financial quality, moat, risks, valuation); selects 3 finalists based on portfolio complementarity; and produces 800-1200-word deep analyses from four master perspectives (Duan on business essence, Buffett on moat, Munger on risks, Li on civilizational trend) with a final report containing recommended portfolio, ETF alternatives, industry position assessment, and A/B/C information sufficiency self-rating. It requires citing data sources and runs 15% data-spot-check with Python tools before release.
- Value investors wanting to systematically shortlist 3 companies in a sector (e.g., AI compute) without being swayed by hype.
- Analysts needing a fast way to filter out 'story stocks' and financially weak companies, retaining only moat-worthy names.
- Individual investors comparing multiple companies in an industry with consistent standards and traceable screening reasons.
- Researchers requiring a report with elimination reasons, sources, and confidence ratings for decision-making.
- AI Agent users wanting a structured process for investment research instead of vague answers.
What are this skill's strengths and limitations?
- Clear funnel with explicit pass/reject criteria at each layer; elimination reasons are traceable, not black-box.
- Multi-market coverage (A-shares, HK, US, unlisted) avoids missing important names.
- Embeds financial rigor tools (e.g., market-cap verification, multi-source cross-validation) to reduce data errors.
- Four-master perspectives create genuine adversarial tension, avoiding single-view blind spots.
- Works on both Claude Code and Codex platforms.
- Backed by a framework with real trading performance (README shows 2024/2025 returns over 66%).
- Heavyweight: requires multiple research rounds and step-by-step screening, high token usage, not for quick simple decisions.
- Relies on external data sources (e.g., Tonghuashun, Futu, Yahoo Finance) but no explicit API; actual use may need manual or web search.
- No test suite or automated tests; report_audit.py exists but coverage is unspecified.
- Subjective judgments on 'moat rating' and 'portfolio complementarity' may be inconsistent across runs.
- China/Asia data may be limited due to fewer English resources, requiring extra effort.
- Output is in Chinese (per skill spec); English users need to translate/adjust.
How do you install this skill?
Install the AI Berkshire repo: git clone https://github.com/xbtlin/ai-berkshire.git, then install skills for your client. For Claude Code: run ./scripts/install-claude-commands.sh (macOS/Linux) or install-claude-commands.bat (Windows) to copy skills to the global commands dir. For Codex: run ./scripts/install-codex-skills.sh (macOS/Linux) or install-codex-skills.bat (Windows) to generate skills to ~/.codex/skills. Restart your client after installation.
How do you use this skill?
In Claude Code: call /industry-funnel AI compute. In Codex: type 'Use industry-funnel to screen AI compute'. The skill reads $ARGUMENTS as the industry, auto-confirms the date, and runs the screening funnel. Report is saved to reports/{industry}-funnel-{YYYYMMDD}.md. Example: 'AI compute' generates four sub-sector logic chains and final 3 picks.
How does this skill compare with similar options?
Compared to the sibling skill industry-research in the same repo: industry-research focuses on industry chain structure and segment slices, while industry-funnel focuses on individual stock screening funnel from whole market to 3 picks; they are complementary.