# How the AI Berkshire Earnings Review Skill Works: Complete Technical Breakdown

> Discover how the AI Berkshire Earnings Review skill works. Explore its eight stage pipeline for deep-read earnings reports with automated data validation and tone analysis.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: deep-dive
- Published: 2026-07-27

---

**The Earnings Review skill is a single‑agent workflow defined in `skills/earnings‑review.md` that executes an eight‑stage pipeline to produce deep‑read earnings reports from primary source documents, including automated data validation and management tone analysis.**

The **Earnings Review skill** in the [AI Berkshire](https://github.com/xbtlin/ai-berkshire) repository transforms raw SEC filings and earnings materials into structured investment research through a reproducible, markdown‑driven pipeline. Unlike generic financial analysis tools, this skill enforces a "read the original filing" philosophy by parallelizing document acquisition, cross‑validating figures against canonical sources, and auditing every datum before publication.

## Eight-Stage Workflow Architecture

The skill operates as a sequential workflow where each stage is implemented as plain‑text instructions in the skill definition file. When invoked via slash‑command (e.g., `/earnings‑review 腾讯 2025Q4`), the AI Berkshire runtime expands `$ARGUMENTS` and executes the following stages:

### Stage 1: Pre‑Screening for Material Availability

Before processing begins, the skill classifies source availability into **Grade A**, **B**, or **C** according to `skills/earnings‑review.md` (lines 26‑31). This rating determines how much original documentation can be used versus reliance on third‑party summaries.

- **Grade A**: Full original filing obtained (10‑K/10‑Q/年报)
- **Grade B**: Partial original text or third‑party aggregation only
- **Grade C**: News reports and data‑site abstracts only

### Stage 2: Parallel Document Acquisition

The skill fires parallel **Task** agents to download four categories of primary materials simultaneously (lines 34‑40):

1. The official filing itself (10‑K/10‑Q/annual report)
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)

### Stage 3: Financial Extraction and Cross‑Validation

Core financial data is parsed from income statements, cash‑flow statements, and balance sheets. The skill then invokes **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)** (lines 78‑92) to perform cross‑validation checks on revenue, market capitalization, and valuation metrics against approved sources listed in [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md).

### Stage 4: Management Tone and MD&A Analysis

The skill conducts line‑by‑line reading of Management Discussion & Analysis (MD&A) sections and earnings call transcripts, flagging five specific signal types (lines 98‑108):

- **🟢 坦诚信号** (Candor signals) – Direct acknowledgment of challenges
- **🔵 清晰信号** (Clarity signals) – Precise, quantified guidance
- **🟡 模糊信号** (Vagueness signals) – Non‑specific optimistic language
- **🟠 转移信号** (Deflection signals) – Shifting blame or topic
- **🔴 归因外部化** (Externalization signals) – Attributing results solely to macro factors

### Stage 5: Footnote and Hidden Signal Detection

A dedicated "note‑digging" phase scans financial statement footnotes for six critical risk categories (lines 31‑38):

- Related‑party transactions (关联交易)
- Equity incentive dilution (股权激励)
- Contingent liabilities (或有负债)
- Accounting policy changes (会计政策变更)
- Segment information splits (分部信息)
- Customer/supplier concentration (客户/供应商集中度)

### Stage 6: Historical Context and Guidance Comparison

The skill constructs multi‑period tables placing current metrics within at least **four quarters** or **three annual reports** of historical context (lines 52‑66). It compares actual results against prior management guidance and records deviations for trend analysis.

### Stage 7: Structured Report Generation

Output is rendered as a structured markdown document with seven mandatory sections (lines 70‑80):

1. 核心数据速览 (Core data snapshot)
2. 本期最重要的 3 个变化 (Top 3 changes this period)
3. 管理层语气与承诺追踪 (Management tone & commitment tracking)
4. 附注中的隐藏信息 (Hidden information in notes)
5. Q&A 关键问题 (Key Q&A highlights)
6. 对投资主题的影响 (Impact on investment thesis)
7. 结论 (Conclusion with hold/add/reduce recommendation)

### Stage 8: Automated Audit and Quality Gate

Before publication, **[`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py)** executes a final "准出" (exit) audit (lines 94‑108). This tool:

- Extracts a data checklist from the generated report
- Re‑fetches every datum from canonical sources in [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md)
- Issues a **pass/fail verdict**; only passing reports are written to disk

## Key Implementation Files and Tools

The Earnings Review skill relies on a specific file architecture within the AI Berkshire repository:

| File | Purpose |
|------|---------|
| `skills/earnings‑review.md` | Master skill definition containing all eight workflow stages |
| [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py) | Validation utility for revenue, market‑cap, and valuation cross‑checks |
| [`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py) | Automated quality gate that re‑fetches and verifies every reported figure |
| [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md) | Canonical catalogue of approved data sources (macrotrends, StockAnalysis, 东方财富) |
| `skills/earnings‑team.md` | Multi‑agent orchestrator that can invoke the single‑agent Earnings Review as a sub‑task |

## Practical Usage and Output Format

Users interact with the skill through natural slash commands. The runtime automatically parses company names and fiscal periods from `$ARGUMENTS`.

**Quarterly analysis with explicit period:**

```text
/earnings-review 腾讯 2025Q4

```

**Annual report auto‑detection (defaults to most recent filing):**

```text
/earnings-review 美团

```

### Expected Output Structure

The generated markdown report follows a standardized template. Below is a truncated example showing the data tables and signal classifications:

```markdown

## 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% | ✅ |
| 经营现金流 / 净利润 | 112% | 98% | +14% | >100% | ⚠️ |

## 二、本期最重要的 3 个变化

1. 毛利润率回落 0.8%——管理层归因于 X 业务的高额研发开支。
2. 资本支出 ↑ 22% → 新增 5 G 基站。
3. 关联交易：与关联方 Y 的采购价格上调 12%。

## 三、管理层语气与承诺追踪

| 信号 | 示例 | 评估 |
|------|------|------|
| 🟢 坦诚信号 | "本季度利润率下降主要因为我们在 X 领域的投入超出预期" | 正向 |
| 🔴 模糊信号 | "我们对未来充满信心" | 负向 |

## 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。
**下一个催化剂** – 8‑月新一代芯片发布。

```

Reports are written to `reports/{company}-earnings-{period}.md` upon passing the audit gate.

## Summary

- The **Earnings Review skill** is defined entirely in `skills/earnings‑review.md` as a reproducible, eight‑stage markdown workflow.
- It enforces primary‑source research by parallelizing acquisition of filings, transcripts, letters, and decks, then rating material availability as Grade A, B, or C.
- Financial rigor is ensured through **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)** validation and a final audit via **[`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py)** against canonical sources in [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md).
- Management communications are analyzed for five specific behavioral signals (candor, clarity, vagueness, deflection, externalization) to assess execution risk.
- The skill outputs a structured seven‑section markdown report suitable for direct inclusion in investment memos.

## Frequently Asked Questions

### What file defines the Earnings Review skill workflow?

The complete workflow is defined in **`skills/earnings‑review.md`** at the repository root. This markdown file contains the eight‑stage instruction set interpreted by the AI Berkshire runtime, including parallel Task agent invocations, signal classification rules, and report templates.

### How does the skill validate financial data accuracy?

During Stage 3, the skill calls **[`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py)** to cross‑validate extracted figures. In Stage 8, **[`tools/report_audit.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/report_audit.py)** extracts a checklist of every numerical claim, re‑fetches each value from approved sources catalogued in [`skills/financial-data.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/financial-data.md), and issues a binary pass/fail verdict before the report is published.

### What are the "signal" categories used in management tone analysis?

The skill identifies five signals in MD&A and earnings call transcripts: **坦诚信号** (candor), **清晰信号** (clarity), **模糊信号** (vagueness), **转移信号** (deflection), and **归因外部化** (externalization). These are tagged line‑by‑line to flag execution risk or governance concerns that raw financial metrics might obscure.

### Can the Earnings Review skill be used as part of a multi-agent workflow?

Yes. While the skill functions as a single‑agent workflow for standard requests, **`skills/earnings‑team.md`** acts as a higher‑level orchestrator that can invoke the Earnings Review skill as a sub‑task within larger multi‑agent deep‑dives, allowing coordinated analysis across multiple companies or fiscal periods.