How to Conduct Earnings Reviews Using AI Berkshire's `/earnings-review` Skill

TLDR: The /earnings-review skill in xbtlin/ai-berkshire executes an eight-step workflow defined in skills/earnings-review.md to automate earnings analysis, using parallel agents for data acquisition and tools/financial_rigor.py for zero-dependency validation before generating audited Markdown reports.

The xbtlin/ai-berkshire repository provides a sophisticated AI-driven framework for financial analysis. The /earnings-review skill automates the entire lifecycle of earnings report analysis, from primary source acquisition to final audit verification. This guide explains how to leverage the skill's eight-step workflow to produce institutional-grade earnings reviews.

Invoking the /earnings-review Skill

To start an earnings review, invoke the skill with the company name and reporting period. The skill automatically resolves company identifiers and initiates the workflow defined in skills/earnings-review.md.

/earnings-review 腾讯 2025Q4

This creates reports/腾讯-earnings-2025Q4.md after completing all eight steps.

The Eight-Step Workflow

The skill orchestrates a full-cycle analysis through eight logical phases. Each step is implemented via specific sections in skills/earnings-review.md and supported by specialized tools.

Step 1: Data Availability Rating (Lines 24-31)

The workflow begins by classifying source availability as Grade A, B, or C (资料可得性评级). This classification sets downstream expectations for data completeness and determines which validation thresholds apply to the current review.

Step 2: Parallel Primary Source Acquisition (Lines 32-41)

Parallel agents simultaneously fetch raw materials including 10-K/10-Q filings, conference call transcripts, and shareholder letters (获取一手资料). The skill spawns multiple background agents via the Task tool to query SEC EDGAR, 港交所披露易, and 巨潮资讯 concurrently, minimizing latency through parallel execution.

Step 3: Financial Data Extraction and Validation (Lines 43-78)

The skill constructs standardized data tables for income statements, cash flow, and balance sheets (核心财务数据提取与验证). It then invokes tools/financial_rigor.py to perform cross-source checks. This zero-dependency validation tool verifies market cap, valuation metrics, and mathematical exactness without external library dependencies.

python3 tools/financial_rigor.py cross-validate \
  --field revenue \
  --values '{"年报": 7518, "Yahoo": 7500, "StockAnalysis": 7520}' \
  --unit 亿

Step 4: MD&A and Management Tone Analysis (Lines 94-108)

The skill parses Management Discussion and Analysis (MD&A) sections and Q&A transcripts (管理层讨论精读), tagging tone signals including frankness (坦诚), clarity (清晰), ambiguity (模糊), deflection (转移), and external attribution (归因外部化). These qualitative markers provide insight into management credibility and communication transparency.

Step 5: Footnote and Hidden Information Mining (Lines 127-138)

Anomaly detection algorithms scan for high-risk items buried in financial statement footnotes (附注与隐藏信息挖掘). The system flags related-party transactions (关联交易), equity incentives (股权激励), contingent liabilities (或有负债), and accounting policy changes (会计政策变更) that may materially impact valuation.

Step 6: Historical Comparison and Guidance Analysis (Lines 148-167)

Current period data is inserted into a multi-quarter trend matrix (与历史数据对比). The skill compares actuals against previous management guidance and historical seasonal patterns to identify deviations and trend anomalies across multiple reporting periods.

Step 7: Structured Report Generation (Lines 170-188)

The skill populates a fixed Markdown template at reports/{company}-earnings-{period}.md (输出精读报告). The template uses a deliberately simple structure to enable downstream skills to reference the output via {{ include }} macros without re-parsing raw data.

Step 8: Audit Pipeline and Quality Gates (Lines 190-208)

Before publication, tools/report_audit.py extracts a verification checklist and re-queries every data point (数据抽检). The audit pipeline returns a publish or reject verdict based on data provenance verification, enforcing an audit-first culture before any report is considered final.

# Extract checklist from generated report

python3 tools/report_audit.py extract \
  --report reports/腾讯-earnings-2025Q4.md

# Verify specific line items (example for revenue)

python3 tools/financial_rigor.py cross-validate \
  --field revenue \
  --values '{"公司财报": 7518, "Yahoo Finance": 7515}' \
  --unit 亿

# Produce final verdict based on verification results

python3 tools/report_audit.py verdict \
  --results '<JSON-payload-from-previous-step>' \
  --report reports/腾讯-earnings-2025Q4.md

Reusing Earnings Reports in Downstream Skills

Generated reports support modular inclusion through the Codex-prompt compatibility layer. Other skills can embed verified analysis using the include macro, which is kept in sync by scripts/sync-codex-prompts.py.

> See the latest earnings insight:
> ```markdown
> {{ include "reports/腾讯-earnings-2025Q4.md" }}
> ```

This allows portfolio-review and thesis-tracker skills to consume verified data without re-parsing raw filings.

Summary

• The /earnings-review skill in xbtlin/ai-berkshire automates eight distinct phases of earnings analysis via skills/earnings-review.md.
• Parallel data harvesting accelerates acquisition from multiple regulatory sources simultaneously through background agents.
• Zero-dependency validation through tools/financial_rigor.py ensures mathematical accuracy and cross-source verification without external library risks.
• The audit-first culture enforced by tools/report_audit.py guarantees data provenance before any report is published at reports/{company}-earnings-{period}.md.
• Output templates use standard Markdown syntax to ensure interoperability with other AI Berkshire skills via the Codex-prompt compatibility layer.

Frequently Asked Questions

What file formats does the earnings review skill generate?

The skill generates Markdown files following the template pattern reports/{company}-earnings-{period}.md. This format is chosen for human readability and machine parsability, allowing other skills to reference outputs via include macros without additional parsing logic.

How does the validation ensure data accuracy?

Accuracy is enforced through tools/financial_rigor.py, which performs cross-source validation, Benford's Law analysis, and exact mathematical calculations using pure Python standard libraries. Before final output, tools/report_audit.py re-queries every data point from reliable sources to verify provenance and returns a binary verdict.

Can I customize the earnings review workflow?

Yes. The workflow is defined in the human-readable skills/earnings-review.md file. Analysts can modify the eight-step sequence, adjust validation thresholds in financial_rigor.py, or customize the Markdown template structure located at lines 170-188 of the skill file to match specific reporting requirements.

Which data sources does the skill query for primary documents?

The skill queries SEC EDGAR for US listings, 港交所披露易 for Hong Kong stocks, and 巨潮资讯 for A-share companies. These queries execute in parallel via background agents to minimize latency during the acquisition phase defined in skills/earnings-review.md lines 32-41.

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