How the `/earnings-team` Skill Converts Quarterly Earnings Analysis into Publish-Ready Articles

The /earnings-team skill deploys a four-stage agentic pipeline that parallelizes research across four master agents, synthesizes consensus-driven analysis, and enforces automated editorial and audit checks to deliver WeChat-ready public articles from raw quarterly earnings data.

The xbtlin/ai-berkshire repository implements a sophisticated multi-agent system for financial content generation. The /earnings-team skill transforms unstructured quarterly earnings filings into rigorously validated public articles through systematic collaboration between specialized research agents and automated quality assurance tools.

Stage 1: Parallel Research by Four Master Agents

As defined in skills/earnings-team.md (lines 55-57), the skill launches four parallel Agent instances simultaneously, each embodying a distinct investment philosophy:

• 段永平: Analyzes revenue structure and competitive positioning
• 巴菲特: Validates cash flow patterns and capital allocation
• 芒格: Examines industry trends and competitive moats
• 李录: Assesses management tone and risk signals

Each agent executes a domain-specific checklist against primary sources including SEC filings, earnings call transcripts, and quantitative datasets. Agents annotate source reliability (A/B/C grades) and validate quantitative claims using the Financial Rigor Toolkit (tools/financial_rigor.py). For example, the 巴菲特 agent invokes cross-source validation (lines 110-113) via:

python3 ~/ai-berkshire/tools/financial_rigor.py cross-validate \
  --metric "revenue" --values {值1} {值2} --sources "来源1" "来源2"

Stage 2: Team Lead Synthesis

After the research agents complete, the Team Lead consolidates outputs as specified in lines 44-48 of skills/earnings-team.md. This synthesis process:

• Identifies consensus viewpoints across the four masters
• Surfaces contradictions requiring resolution
• Flags omitted analytical angles

The result is a structured markdown research report (template defined in lines 52-84) containing a one-sentence headline, three key changes, a four-master scoring table, core data tables, and actionable investment implications ("what each master would do").

Stage 3: Editorial Production and Review

The pipeline splits into two parallel publication tracks defined in skills/earnings-team.md (lines 91-174):

Editor Agent (Agent 5) rewrites the research report into reader-friendly content applying:

• 倒金字塔 (inverted pyramid) narrative structure
• Concise, WeChat-compatible formatting
• Preservation of all data citations and master quotes

Reader Review Agent (Agent 6) evaluates the draft against four criteria:

• Readability and information density
• Credibility of sources
• Actionable guidance for investors
• Mandatory compliance checks

The Team Lead incorporates mandatory edits from the Reader Review before proceeding to final validation.

Stage 4: Automated Quality Gate and Publication

Before marking the article 【准出】 (approved for release), the skill executes rigorous validation through tools/report_audit.py (lines 14-23 of skills/earnings-team.md):

python3 ~/ai-berkshire/tools/report_audit.py extract \
  --report reports/{公司名}/{公司名}-earnings-{期间}.md

python3 ~/ai-berkshire/tools/report_audit.py verdict \
  --results '<填好的JSON>' \
  --report {报告文件名}

The audit leverages the Financial Rigor Toolkit to perform market capitalization verification, cross-source validation, Benford's law analysis for numerical authenticity, and exact calculation checks. Only when all sanity checks pass does the article receive final publication approval and output to reports/{company}/{company}-earnings-{period}.md.

Key Implementation Files

File	Function
skills/earnings-team.md	Orchestrates the four-stage pipeline and agent coordination (lines 44-174)
tools/financial_rigor.py	Provides command-line validation for financial metrics and cross-source checks
tools/report_audit.py	Performs final extraction and automated quality verification (lines 14-23)
reports/{company}/{company}-earnings-{period}.md	Destination path for publish-ready articles

Summary

• The /earnings-team skill implements a four-stage pipeline (Research, Synthesis, Publication, Quality Gate) to transform raw earnings data into public articles.
• Four parallel master agents (段永平, 巴菲特, 芒格, 李录) conduct specialized quarterly earnings analysis simultaneously, each following domain-specific checklists defined in skills/earnings-team.md (lines 55-57).
• The Team Lead synthesizes multi-perspective analysis into a structured markdown report (lines 52-84), identifying consensus and contradictions.
• Dual publication agents (Editor and Reader Review) enforce reader-centric formatting and quality standards (lines 91-174) before the final audit.
• Automated validation via tools/report_audit.py and tools/financial_rigor.py ensures numerical accuracy and source reliability before marking content 【准出】.

Frequently Asked Questions

What are the four master agents in the research stage?

The research stage deploys four parallel agents embodying distinct investment philosophies: 段永平 (revenue structure analysis), 巴菲特 (cash flow validation), 芒格 (competitive trend assessment), and 李录 (management tone and risk signal detection). Each agent operates according to a specialized checklist defined in skills/earnings-team.md (lines 55-57), pulling from primary sources like SEC filings and earnings call transcripts.

How does the skill ensure numerical accuracy before publication?

Numerical accuracy is enforced through the Financial Rigor Toolkit (tools/financial_rigor.py) and the final Quality Gate. The system runs tools/report_audit.py (lines 14-23) to execute market-cap verification, cross-source validation using the cross-validate function, Benford's law checks for data authenticity, and exact calculation verification. Only articles passing all automated audits receive the 【准出】 designation.

What editorial standards does the publication stage enforce?

The Editor Agent applies 倒金字塔 (inverted pyramid) structure, ensuring critical information appears first with concise headings and WeChat-compatible formatting. The Reader Review Agent subsequently evaluates readability, information value, credibility, and actionable guidance. The Team Lead must resolve all mandatory edits flagged by the Reader Review before the article proceeds to the audit phase.

How does the Team Lead resolve contradictions between agent analyses?

During the Synthesis stage (lines 44-48 of skills/earnings-team.md), the Team Lead programmatically merges the four master outputs to identify consensus viewpoints, highlight contradictions, and detect omitted angles. This structured reconciliation produces a unified research report that acknowledges divergent perspectives while prioritizing data-backed consensus, mirroring human research team deliberation workflows.

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