How the AI Berkshire Earnings Review Skill Works: Complete 8-Stage Technical Breakdown

The AI Berkshire earnings review skill operates as a single-agent markdown-defined workflow that downloads primary financial documents, cross-validates data against approved sources, analyzes management tone signals, and passes through a rigorous audit gate before generating a structured investment report.

The earnings review skill in the xbtlin/ai-berkshire repository implements a "deep-read" philosophy by processing first-hand source material through eight discrete analytical stages. Defined entirely in the Markdown-based skill file skills/earnings-review.md, this agentic workflow transforms raw SEC filings, earnings call transcripts, and footnotes into audited investment analysis without requiring human intervention in the data pipeline.

What Is the AI Berkshire Earnings Review Skill?

The earnings review skill is a single-agent workflow designed to produce institutional-grade earnings reports from primary source materials. Unlike automated scrapers that rely on third-party summaries, this skill orchestrates parallel document acquisition, financial rigor checks, and semantic analysis of management discussions to surface alpha-generating signals hidden in regulatory filings. The runtime interprets the Markdown instructions in skills/earnings-review.md as executable logic, allowing the skill to be invoked via slash commands like /earnings-review 腾讯 2025Q4.

The 8-Stage Operational Pipeline

The skill executes an eight-stage pipeline, with each step codified in the skill definition file and supported by specific utility modules in the tools/ directory.

Stage 1: Material Availability Rating (A/B/C Classification)

Before processing begins, the system classifies source document availability into three tiers:

• A级 (Grade A): Full original documents obtained (10-K/10-Q/年报, earnings transcripts, shareholder letters)
• B级 (Grade B): Partial original documents or third-party compilations only
• C级 (Grade C): News reports and data website abstracts only

This rating determines the depth of original documentation used in subsequent analysis and is recorded at lines 26-31 of skills/earnings-review.md.

Stage 2: Parallel Document Acquisition via Task Agents

The skill fires parallel Task agents to download four categories of primary materials simultaneously:

1. Regulatory filings (10-K/10-Q/年报)
2. Earnings call transcripts (业绩电话会纪要)
3. Management shareholder letters (管理层致股东信)
4. Investor-day presentation decks (投资者日材料)

This parallel acquisition prevents cascading delays if individual sources are slow to respond, as implemented in lines 34-40 of the skill definition.

Stage 3: Financial Data Extraction and Cross-Validation (financial_rigor.py)

After acquisition, the pipeline parses income-statement, cash-flow, and balance-sheet items, then invokes tools/financial_rigor.py to perform cross-validation checks on revenue consistency, market-cap calculations, and valuation metrics. This verification step ensures that extracted figures match across multiple primary sources before entering the analysis phase (lines 78-92).

Stage 4: Management Tone and MD&A Signal Analysis

The skill performs line-by-line semantic analysis of Management Discussion & Analysis (MD&A) sections and earnings call transcripts, flagging five specific signal types:

• 🟢 坦诚信号 (Candid signals): Explicit acknowledgment of problems or misses
• 🟢 清晰信号 (Clear signals): Specific, quantified guidance or explanations
• 🔴 模糊信号 (Vague signals): Generic confidence statements without specifics
• 🔴 转移信号 (Deflecting signals): Blaming external factors or avoiding the question
• 🔴 归因外部化 (External attribution): Excessive blame on macro conditions or competitors

This tonal analysis occurs at lines 98-108 and contributes to the final investment thesis.

Stage 5: Footnote Mining for Hidden Risks

The pipeline automatically inspects financial statement footnotes for six critical risk categories (必查附注项):

• 关联交易 (Related-party transactions): Pricing and volume with affiliates
• 股权激励 (Equity incentives): RSU/option dilution impacts
• 或有负债 (Contingent liabilities): Off-balance-sheet obligations
• 会计政策变更 (Accounting policy changes): Revenue recognition or depreciation method shifts
• 分部信息 (Segment splits): Geographic or product-line performance dispersion
• 客户/供应商集中度 (Customer/supplier concentration): Revenue dependency risks

This stage, defined at lines 31-38, often uncovers material information absent from headline figures.

Stage 6: Historical Trend and Guidance Deviation Analysis

The system constructs multi-period comparison tables spanning at least 4 quarters (or 3 annual reports) to contextualize current performance. It specifically compares actual results against prior management guidance (管理层指引), recording deviations and accounting for any guidance revisions made during the period (lines 52-66).

Stage 7: Structured Markdown Report Assembly

Verified data and analytical signals are rendered into a standardized markdown structure containing:

1. 核心数据速览 (Core data snapshot): Key metrics with YoY and QoQ deltas
2. 本期最重要的 3 个变化 (Top 3 changes): Material business developments
3. 管理层语气与承诺追踪 (Tone & commitment tracking): Signal inventory with examples
4. 附注中的隐藏信息 (Hidden signals in notes): Footnote-derived risks
5. 历史对比与预期差 (Historical variance): Guidance tracking tables
6. 问答质量 (Q&A quality): Analyst question depth assessment
7. 结论 (Conclusion): Investment thesis impact and position recommendation

The report is written to the reports/ directory with naming convention {company}-earnings-{period}.md.

Stage 8: Automated Audit and Quality Gate (report_audit.py)

Before publication, the skill triggers tools/report_audit.py to execute a three-step verification:

1. Extract: Auto-generates a data checklist from the report content
2. Re-fetch: Retrieves every cited datum from approved sources listed in skills/financial-data.md (macrotrends, StockAnalysis, 东方财富, etc.)
3. Verdict: Produces a pass/fail determination based on data consistency

Only reports passing this audit gate are finalized, ensuring zero unverified figures reach the output (lines 94-108).

Invoking the Skill: CLI Usage and Examples

Users interact with the skill through slash commands. The runtime automatically expands $ARGUMENTS into the user-provided company ticker and reporting period.

# Quarterly review with explicit period

/earnings-review 腾讯 2025Q4

When invoked, the skill executes all eight stages and outputs a structured analysis:

## 一、核心数据速览

| 指标 | 本期 | 上期 | YoY | 管理层指引 | 是否达标 |
|------|------|------|-----|------------|----------|
| 总收入 | 527.3 B | 508.1 B | +3.8% | 530 B | ✅ |
| 毛利润率 | 45.2% | 46.0% | -0.8% | ≥45% | ✅ |

## 四、附注中的隐藏信息

- 关联交易：关联方 Y 关联交易金额 3.2 B，定价高于行业平均 15%
- 股权激励：本期 RSU 发行 1.1 M 股，稀释率 0.6%

## 七、结论

**超预期** – 收入与毛利率基本符合指引，现金流质量提升  
**对持仓** – 维持/适度增持

For annual report reviews without specifying a period, the skill auto-detects the most recent filing:

# Auto-detects latest annual report (年报)

/earnings-review 美团

Core Architecture and File Dependencies

The earnings review skill relies on a specific file ecosystem within xbtlin/ai-berkshire:

File	Purpose
skills/earnings-review.md	Primary skill definition containing all 8 stage instructions
tools/financial_rigor.py	Validation engine for revenue, market-cap, and valuation sanity checks
tools/report_audit.py	Quality gate implementation (extract and verdict commands)
skills/financial-data.md	Canonical source catalog (macrotrends, StockAnalysis, 东方财富) for audit re-fetch
codex-skills/earnings-review/SKILL.md	Codex-compatible wrapper auto-generated from the main skill
skills/earnings-team.md	Multi-agent orchestrator that can invoke this skill as a sub-task

Summary

• The earnings review skill is defined in skills/earnings-review.md as an 8-stage markdown workflow
• It implements a material availability rating (A/B/C) before processing to ensure source quality
• Parallel Task agents acquire primary documents while financial_rigor.py validates numerical integrity
• Semantic analysis flags five distinct management tone signals (candid, clear, vague, deflecting, externalizing)
• Automated footnote mining detects related-party risks, dilution, and contingent liabilities
• The report_audit.py utility enforces a zero-tolerance data verification gate before publication
• Output follows a strict 7-section markdown template suitable for investment committee review

Frequently Asked Questions

How does the earnings review skill verify data accuracy?

The skill employs a two-layer verification system. During Stage 3, tools/financial_rigor.py cross-validates extracted figures across income statements, cash flows, and balance sheets. During Stage 8, tools/report_audit.py re-fetches every cited data point from approved sources cataloged in skills/financial-data.md (such as macrotrends and 东方财富), comparing re-fetched values against the report content to ensure zero errors before publication.

What is the difference between Grade A, B, and C material ratings?

Grade A (A级) indicates the system obtained complete original filings and transcripts, enabling deep fundamental analysis. Grade B (B级) means only partial originals or third-party summaries were available, limiting footnote analysis. Grade C (C级) signifies reliance on news reports only, triggering a caution flag in the final report regarding source reliability.

Can the earnings review skill analyze companies outside the US?

Yes. The skill supports both SEC filings (10-K/10-Q) and Chinese regulatory filings (年报). The parallel document acquisition stage specifically searches for localized primary materials including Chinese earnings call transcripts and investor-day decks, while skills/financial-data.md maintains approved data sources for both US and Chinese equities.

How does the skill detect manipulative management language?

During Stage 4, the skill scans MD&A and call transcripts for linguistic patterns associated with opacity. It specifically flags 模糊信号 (vague signals) like generic confidence statements lacking specifics, 转移信号 (deflecting signals) where questions are evaded, and 归因外部化 (external attribution) where management excessively blames macro factors. These signals are tabulated with direct quotes to support the final tone assessment.

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