# How AI Berkshire Performs Deep Dives into Company Management: A Technical Breakdown of the /management-deep-dive Skill > Explore how AI Berkshire performs management deep dives. Discover the technical breakdown of the /management-deep-dive skill and its four parallel agents for robust company analysis. - Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire) - Tags: how-to-guide - Published: 2026-06-27 --- **AI Berkshire conducts management deep dives through a dedicated `/management-deep-dive` skill that deploys four parallel agents to independently assess CEO ability, integrity, capital allocation discipline, and governance structures, culminating in a weighted composite score derived from historical data and side‑channel validation.** The **xbtlin/ai-berkshire** repository implements Buffett‑style management analysis as a reproducible, automated workflow. At the heart of this capability lies the **`/management-deep-dive`** skill, defined in [`skills/management-deep-dive.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/management-deep-dive.md) and auto‑generated for Codex environments in [`codex-skills/management-deep-dive/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/management-deep-dive/SKILL.md). This skill follows a three‑layer architecture—Skill → Agent → Tool—ensuring that subjective management quality is converted into quantifiable, evidence‑based metrics. ## The Three‑Layer Architecture Behind Management Analysis The deep‑dive skill adheres to the repository’s **"Three‑Layer Design Philosophy"** documented in the README and visualized in `assets/architecture.mmd`. - **Skill Layer**: Defines the command interface (`/management-deep-dive`) and orchestration logic. - **Agent Layer**: Spawns concurrent background agents to gather heterogeneous data streams. - **Tool Layer**: Provides precise financial calculations via utilities like [`tools/financial_rigor.py`](https://github.com/xbtlin/ai-berkshire/blob/main/tools/financial_rigor.py). This separation ensures that management assessment remains deterministic while leveraging parallel processing for speed. ## Parallel Agent Workflow for Data Collection Upon invocation, the skill executes a **layered, parallel‑agent architecture** (lines 53‑58 of [`management-deep-dive.md`](https://github.com/xbtlin/ai-berkshire/blob/main/management-deep-dive.md)). First, a web‑search step identifies key executives (CEO, CFO, founder, controlling shareholder). Then, four agents run simultaneously: 1. **Agent 1** – Scrapes CEO public statements, shareholder letters, earnings calls, and social media to capture strategic forecasts and communication patterns. 2. **Agent 2** – Tracks capital‑allocation actions including M&A history, buy‑back programs, dividend policies, and new‑business investments. 3. **Agent 3** – Analyzes governance mechanisms, examining equity structures (AB‑shares, super‑voting rights, VIE arrangements), executive compensation schemes, and related‑party transactions. 4. **Agent 4** – Performs side‑channel validation by aggregating employee reviews (Glassdoor), customer feedback (App Store ratings), and industry reputation (Zhihu, professional forums). Each agent operates independently, ensuring that biases in one data stream do not contaminate others. ## CEO Ability‑Circle Assessment The skill evaluates leadership through the **Ability‑Circle** framework (sections 2.1 and 2.2). - **Strategic Foresight**: Measures the accuracy of historical market predictions and long‑term vision statements against actual outcomes. - **Execution Capability**: Assesses delivery on promises, talent attraction metrics, and crisis‑management responses documented in public records. This dual‑lens approach distinguishes between leaders who articulate vision and those who consistently deliver results. ## Integrity Evaluation via Commitment Fulfillment Integrity assessment quantifies the **commit‑vs‑delivery** gap across a rolling three‑year window (sections 3.1 and 3.2). The skill calculates a **commitment fulfillment rate** by comparing public pledges (earnings guidance, strategic targets) with subsequent financial reports and operational milestones. Additionally, it analyzes leadership behavior during crises—such as capital allocation decisions during downturns—to infer ethical consistency and shareholder alignment. ## Capital‑Allocation Rigor Analysis Every major capital decision receives a quantitative score (section 4.1 and 4.2). The skill logs: - **M&A transactions** and their subsequent ROIC performance - **Buy‑back programs** and their timing relative to intrinsic value - **Dividend sustainability** based on free‑cash‑flow coverage - **New‑business investments** and their strategic fit Critically, these decisions are cross‑checked with **`tools/financial_rigor.py verify-valuation`**, a utility that validates whether capital deployments occurred at reasonable valuations relative to intrinsic value estimates. ## Governance Structure Assessment The skill examines governance quality through three lenses (sections 5.1‑5.3): - **Equity Architecture**: Identifies dual‑class structures, super‑voting rights, and founder control premiums that insulate management from shareholder feedback. - **Compensation Rationality**: Evaluates whether pay structures align long‑term incentives with sustainable value creation rather than short‑term stock price manipulation. - **Related‑Party Transactions**: Flags potential conflicts of interest and tunneling risks through historical transaction analysis. ## Side‑Channel Validation and Sentiment Triangulation To mitigate reliance on curated corporate communications, the skill incorporates **side‑channel validation** (sections 6.1‑6.3). This phase scrapes unstructured data sources—employee morale indicators, customer satisfaction scores, and industry expert opinions—to triangulate the management’s reputation. Discrepancies between internal narratives and external sentiment trigger deeper investigation into potential cultural or operational risks. ## CEO Succession Risk Scenario The workflow includes a **succession stress test** (section 7). The skill asks: *"What happens if the CEO leaves tomorrow?"* It evaluates organizational depth, the existence of formal succession plans, and whether competitive advantages depend on the founder’s personal network versus institutionalized processes. This addresses the key‑person dependency risk central to Buffett’s "buy the people, not just the stock" philosophy. ## Report Generation and Scoring Methodology Findings compile into a structured markdown report saved at `reports/{公司名}-management-{YYYYMMDD}.md` (sections 8‑9). The report contains: - Key personnel inventory and tenure analysis - Integrity score with historical commitment fulfillment trends - Ability‑Circle ratings with specific evidence citations - Governance risk flags - Side‑channel sentiment summaries - **Weighted Composite Score**: - **Integrity**: 35% - **Ability**: 25% - **Capital‑Allocation**: 25% - **Governance**: 15% This weighting reflects the Buffett maxim that integrity outweighs raw intelligence in management quality. ## Invoking the Management Deep‑Dive Skill The skill supports both direct command‑line invocation and natural‑language interaction in Codex environments. **Standard CLI usage:** ```bash # Analyze a listed company /management-deep-dive 美团 # Analyze a specific executive within a company /management-deep-dive 王兴 美团 # Analyze an unlisted private firm /management-deep-dive SpaceX ``` **Codex environment syntax:** ```text 使用 management-deep-dive 研究 美团 使用 management-deep-dive "王兴 美团" 使用 management-deep-dive SpaceX ``` ## Summary - **AI Berkshire** implements management analysis as the **`/management-deep-dive`** skill, residing in [`skills/management-deep-dive.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/management-deep-dive.md) and [`codex-skills/management-deep-dive/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/management-deep-dive/SKILL.md). - The architecture employs **four parallel agents** to simultaneously gather data on CEO statements, capital allocation, governance, and external sentiment. - **Integrity** (35%) receives the highest weight in the final score, followed by **Ability** (25%), **Capital‑Allocation** (25%), and **Governance** (15%). - Financial rigor is enforced through **`tools/financial_rigor.py verify-valuation`**, which validates capital deployment decisions against intrinsic value. - Output generates as a timestamped markdown report at `reports/{company}-management-{YYYYMMDD}.md` with actionable scoring and risk flags. ## Frequently Asked Questions ### How does the skill handle private companies without public filings? The skill adapts its data collection strategy for unlisted firms by weighting **Agent 4** (side‑channel validation) more heavily, utilizing employee reviews, customer feedback, and industry interviews while acknowledging data limitations in the final report’s confidence interval. ### What is the "commitment fulfillment rate" and how is it calculated? The **commitment fulfillment rate** is a three‑year rolling metric that compares public management commitments—such as earnings guidance or strategic targets—with subsequent actual performance. The skill quantifies this as a percentage of promises kept, with adjustments for materiality and external market shocks. ### Can the skill analyze individual executives rather than entire companies? Yes. By passing the executive name followed by the company, such as `/management-deep-dive 王兴 美团`, the skill focuses its **Agent 1** and **Agent 3** analysis on that specific individual’s track record, public statements, and compensation history while maintaining the standard governance and side‑channel checks. ### How does the three‑layer architecture ensure analysis consistency? The **Skill Layer** standardizes the workflow definition, the **Agent Layer** ensures parallel data collection reduces single‑source bias, and the **Tool Layer** provides deterministic financial calculations (e.g., `verify-valuation`). This separation prevents hallucination by grounding qualitative assessments in quantitative verification tools.