# How the Quality-Screen Skill Applies Seven Hard Filters for Stock Screening in AI-Berkshire > Discover how the quality-screen skill in XBTlin AI Berkshires uses seven hard filters like ROE, FCF, and margins for effective stock screening. Learn how to identify top stocks quickly. - Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire) - Tags: how-to-guide - Published: 2026-07-03 --- **The quality-screen skill in `xbtlin/ai-berkshire` implements a "go-no-go" workflow that evaluates stocks against seven strict financial thresholds—ROE ≥8%, positive 5-year FCF, interest coverage ≥2x, gross margin ≥15%, OCF/NI ratio ≥0.7, net margin ≥5%, and share dilution ≤20%—while allowing specific exemptions for strategic growth cases.** The **quality-screen** skill is a Claude-Code/Codex-compatible workflow designed for disciplined equity screening. Located in [`skills/quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/quality-screen.md), this tool automates a rigorous "eliminate the inferior" (去劣) process that parses ticker lists, collects financial data in parallel, and enforces hard cut-offs to filter out low-quality companies before they reach deeper analysis. ## The Seven Hard Filters Defined in quality-screen.md The core screening logic resides in [`skills/quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/quality-screen.md) (lines 24–35), where seven indicators serve as non-negotiable gates. A company failing any filter is excluded unless it qualifies for a specific exemption. ### Capital Efficiency and Cash Reality Metrics **Filter 1: 10-Year Average ROE ≥ 8%** The skill requires a minimum **10-year average Return on Equity (ROE)** of 8%. This threshold ensures shareholders' capital generates returns above the long-term opportunity cost. In [`quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/quality-screen.md), this filter appears first in the hard-filter table, emphasizing capital efficiency as the primary quality indicator. **Filter 2: 5-Year Cumulative Free Cash Flow (FCF) Must Be Positive** The skill rejects companies with negative cumulative FCF over five years. This captures the "cash reality" check—verifying that profits are backed by actual cash generation rather than accounting adjustments. ### Profitability and Financial Health Thresholds **Filter 3: Interest Coverage Ratio ≥ 2×** Calculated as EBIT divided by interest expense, this ratio must exceed 2× to confirm debt serviceability. The source code explicitly notes that **banks and insurance companies are exempt** from this filter due to their interest-bearing business models. **Filter 4: Long-Term Gross Margin ≥ 15%** This threshold tests pricing power. Companies with gross margins below 15% typically lack product differentiation or competitive moats, making them vulnerable to price wars. **Filter 5: Operating Cash Flow to Net Income Ratio ≥ 0.7** Measured as a 5-year average, this ratio validates earnings quality. If operating cash flow consistently falls below 70% of net income, the skill flags potential accrual accounting distortions or working capital inefficiencies. **Filter 6: Long-Term Net Margin ≥ 5%** The skill requires sustainable net margins above 5% to ensure the company can withstand revenue shocks without immediate losses. This serves as a resilience buffer against cyclical downturns. ### Shareholder Protection Criteria **Filter 7: 5-Year Share Dilution ≤ 20%** Excluding M&A-related issuance, total share dilution over five years must not exceed 20%. This protects existing shareholders from excessive equity financing that erodes ownership value without proportional value creation. ## How the Screening Workflow Executes The skill implements a modular pipeline defined in [`skills/quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/quality-screen.md), separating data ingestion from rule evaluation to ensure reproducibility. ### Input Parsing and Mode Detection In lines 5–12 of [`quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/quality-screen.md), the skill accepts a single `$ARGUMENTS` string that triggers three operating modes: - **Individual mode**: Comma-separated tickers (e.g., `"腾讯, 美团"`) - **Industry mode**: Thematic searches (e.g., `"中国啤酒行业"`) - **Index mode**: Benchmark constituents (e.g., `"沪深300"`) For batch operations, the skill automatically expands industry or index inputs by querying constituent companies via web search before proceeding to data collection. ### Parallel Data Collection via Background Agents Lines 81–94 of the source file describe the **parallel agent architecture**. For each target company, the skill spawns a dedicated background Agent that simultaneously queries: - Annual reports (primary source) - Broker research (secondary validation) - Financial data platforms (real-time verification) Each agent retrieves the seven data points required for the hard filters (ROE, FCF, interest coverage, gross margin, OCF/NI, net margin, and share dilution), enabling concurrent processing that reduces total runtime linearly with the number of target companies. ### Filter Evaluation and Exemption Logic Once data is collected, the skill evaluates each company against the seven thresholds (lines 24–35). If a company fails a specific filter, the workflow checks three exemption rules (lines 36–63) before finalizing exclusion: 1. **Strategic-Investment Period**: Applies to young, high-margin companies (listed <10 years, gross margin >30%, cash-flow positive last 2 years) that may have ROE below 8% due to growth investments. 2. **Intentional Low-Profit**: Covers high-margin firms reinvesting for market share (gross margin >30% with net margin ≥5% in recent years or trending upward). 3. **High-Turnover Thin-Profit**: For membership/platform models (ROE >20%, OCF/NI >1.0, confirmed business type) where margins appear thin but capital efficiency is high. If any exemption criteria are satisfied, the hard filter failure is waived and the company receives a **"通过"** (pass) status. ## Generating Screening Reports After processing all companies, the skill aggregates results into a structured markdown report (lines 96–138). The output includes: - **Pass list**: Companies satisfying all seven filters (or qualifying exemptions) - **Exclusion list**: Companies failing filters without valid exemptions - **Boundary cases**: Companies with data gaps or marginal failures requiring manual review - **Sector insights**: Aggregate statistics showing pass rates by industry The report follows a strict markdown template defined in lines 66–89, ensuring compatibility with downstream automation tools like Slack bots or email pipelines. ## Practical Usage Examples ### Screening Individual Tickers ```python # Single-ticker mode with Chinese equities result = await agent.run( skill="quality-screen", arguments="腾讯, 美团, 英伟达" ) print(result) # Returns markdown report with pass/fail per company ``` ### Industry-Wide Screening ```python # Automatic expansion of beer industry constituents result = await agent.run( skill="quality-screen", arguments="中国啤酒行业" ) ``` ### Index Constituent Screening ```python # Full CSI 300 index screening result = await agent.run( skill="quality-screen", arguments="沪深300" ) ``` ## Summary - **Seven hard filters** in [`skills/quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/quality-screen.md) enforce minimum standards for ROE (≥8%), FCF (positive 5-year), interest coverage (≥2×), gross margin (≥15%), OCF/NI (≥0.7), net margin (≥5%), and dilution (≤20%). - **Parallel agent architecture** spawns background processes per company to fetch financial data from annual reports, broker research, and data platforms simultaneously. - **Three exemption rules** allow strategic cases (young growth companies, reinvestment phases, asset-light platforms) to bypass specific filters when they meet alternative criteria. - **Strict markdown output** standardizes results into pass lists, exclusion lists, and sector insights for automated downstream consumption. ## Frequently Asked Questions ### What happens if a company fails just one of the seven hard filters? According to the logic in [`quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/quality-screen.md) lines 36–63, a failing company is immediately flagged for exclusion unless it satisfies one of the three exemption rules (strategic-investment, intentional low-profit, or high-turnover thin-profit). The skill operates on a "漏网不可误杀" (better to miss a good company than let a bad one through) philosophy, making filters strict but allowing overrides only for specific business model justifications. ### Are financial institutions treated differently by the quality-screen skill? Yes. Lines 52–60 of the source code explicitly exempt banks and insurance companies from the **interest-coverage filter** (Filter 3), as their business models inherently involve interest-bearing liabilities. Additionally, REITs receive special handling for ROE calculations due to their high leverage structures and regulatory distribution requirements. ### How does the skill handle incomplete financial data? The parallel data collection agents (lines 81–94) prioritize sources hierarchically: annual reports first, then broker research, then financial platforms. If critical data points remain missing after exhausting these sources, the skill places the company in a **boundary cases** section of the final report rather than making assumptions, allowing human analysts to decide whether to procure additional data or exclude the security. ### Can the seven filter thresholds be customized without modifying code? Yes. Because the thresholds are stored in a declarative markdown table within [`skills/quality-screen.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/quality-screen.md) (lines 24–35), non-engineers can adjust the cut-off values (e.g., changing ROE from 8% to 10%) by editing the table directly. The [`scripts/sync-codex-skills.py`](https://github.com/xbtlin/ai-berkshire/blob/main/scripts/sync-codex-skills.py) utility ensures these changes propagate to the Codex-compatible prompt version automatically, maintaining synchronization between Claude-Code and Codex environments.