How AI Berkshire Tracks Investment Theses with the /thesis-tracker Skill

The /thesis-tracker skill implements a post-buy discipline system that continuously monitors investment theses by creating structured markdown reports and re-evaluating core assumptions quarterly to generate a numerical Thesis Health Score.

The xbtlin/ai-berkshire repository provides an AI-powered investment analysis framework that transforms static investment ideas into living documents requiring rigorous validation. At the center of this methodology sits the /thesis-tracker skill, which automates the creation and monitoring of investment theses through a dual-mode workflow defined in skills/thesis-tracker.md.

The Two-Mode Workflow Architecture

The skill operates in two distinct modes depending on whether a thesis file already exists for the target company.

Mode A: Establishing a New Investment Thesis

When initiating coverage on a new position, the skill creates a comprehensive thesis document at reports/{Company}-thesis.md. This process begins by gathering current market data and validating valuation metrics through tools/financial_rigor.py.

The generated markdown report contains four critical sections:

• Five-Sentence Investment Thesis – Answering key investment questions in exactly five sentences, establishing the core bullish case【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L44-L53】
• Core Assumption List – Enumerating testable hypotheses with verification methods and monitoring frequencies【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L61-L68】
• Red-Line List – Deal-breaker conditions that trigger immediate re-evaluation or exit【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L71-L78】
• Valuation Anchors – Current price, PE ratio, market cap, intrinsic value estimate, and safety margin calculations【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L85-L91】

Mode B: Tracking Existing Theses

When a thesis file exists, the skill enters monitoring mode, executing a seven-step re-evaluation workflow:

1. Load Historical Data – Retrieves the stored thesis, assumptions, red-lines, and previous check records【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L8-L14】
2. Fetch Live Data – Retrieves latest earnings, news, price movements, and insider-trading data via WebSearch【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L18-L22】
3. Assumption Validation – Re-evaluates each core assumption, updating status indicators in a tracking table【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L24-L33】
4. Red-Line Monitoring – Checks every deal-breaker condition for violation【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L40-L46】
5. Valuation Refresh – Updates valuation anchors with current market data【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L74-L82】
6. Health Score Calculation – Computes the Thesis Health Score using the weighted formula【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L89-L92】
7. Append Results – Records the new check results to the thesis markdown file with timestamp【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/thesis-tracker.md#L94-L100】

Thesis Health Score Calculation methodology

The Thesis Health Score quantifies thesis validity on a numerical scale using the following weighted formula implemented in the skill logic:

health = 10 \
    - (⚫ * 3) \   # broken assumptions

    - (🔴 * 2) \   # damaged assumptions  

    - (🟡 * 1) \   # weakened assumptions

    - (red_lines * 5)

Status indicators follow a traffic-light system:

• 🟢 Healthy: Assumption remains valid
• 🟡 Weakened: Minor deviation requiring observation
• 🔴 Damaged: Significant deterioration threatening thesis
• ⚫ Broken: Invalidated assumption requiring thesis revision

Red-line violations incur a severe 5-point penalty each, reflecting their status as deal-breakers that typically warrant immediate position closure.

Technical Implementation and Source Files

The thesis-tracking capability spans multiple components within the repository:

Component	File Path	Description
Skill Definition	skills/thesis-tracker.md	Canonical workflow documentation and prompt templates
Codex Adapter	codex-skills/thesis-tracker/SKILL.md	Generated adapter mapping markdown to callable skill
Prompt Wrapper	codex-prompts/thesis-tracker.md	Codex entry point prompt configuration
Validation Tool	tools/financial_rigor.py	Utility for verifying valuation calculations
Output Reports	reports/{Company}-thesis.md	Generated living documents containing tracking tables

According to the xbtlin/ai-berkshire source code, the skill automatically invokes WebSearch for data retrieval and executes python3 tools/financial_rigor.py verify-valuation to ensure mathematical rigor in all valuation anchors.

Practical Usage Examples

Execute the thesis-tracker via command line to either establish new coverage or update existing positions:

# Create initial thesis for a new company

ai_berkshire /thesis-tracker "Tesla 建立论文"

# Run quarterly check for an existing position  

ai_berkshire /thesis-tracker "Tesla 季度检查"

The skill interprets the command arguments to determine which mode to execute. When the thesis file exists, it automatically triggers Mode B (tracking); when absent, it initiates Mode A (establishment).

The structured output enables downstream automation, allowing skills such as /thesis-drift or /portfolio-review to consume the health scores and assumption statuses programmatically for broader portfolio management decisions.

Summary

• The /thesis-tracker skill implements a dual-mode workflow that creates new theses or monitors existing ones through structured markdown documents.
• Four core components comprise every thesis: a five-sentence summary, testable assumptions, red-line deal breakers, and valuation anchors.
• Assumption statuses use emoji indicators (🟢🟡🔴⚫) to track validity, feeding into a weighted health score formula that subtracts penalties for weakened, damaged, or broken assumptions.
• Red-line violations automatically subtract 5 points each from the health score, reflecting their severity as thesis-invalidating events.
• The system automatically invokes tools/financial_rigor.py and WebSearch to ensure all valuations and market data remain current and mathematically sound.

Frequently Asked Questions

How does the /thesis-tracker skill determine when to create versus update a thesis?

The skill checks for the existence of a thesis file at reports/{Company}-thesis.md. If the file is absent, it executes Mode A to establish a new thesis by gathering initial data and creating the structured markdown template. If the file exists, it automatically enters Mode B to append a new quarterly check record to the existing document.

What triggers a red-line violation and how does it affect the health score?

Red-lines represent deal-breaker conditions defined during thesis establishment, such as "CEO departure" or "core product revenue decline >20%." Each triggered red-line automatically subtracts 5 points from the 10-point health scale, typically reducing the score to exit territory (≤3 points), which signals immediate position liquidation according to the skill's recommendation logic.

Can the thesis-tracker integrate with other AI Berkshire skills?

Yes. The skill outputs standardized markdown tables containing health scores, assumption statuses, and valuation updates that downstream skills like /thesis-drift (for thesis deviation analysis) and /portfolio-review (for aggregate portfolio health) can parse programmatically. This modular architecture enables comprehensive portfolio management workflows where individual thesis health informs broader allocation decisions.

Where are the thesis documents physically stored in the repository?

Thesis files reside in the reports/ directory using the naming convention {Company}-thesis.md. For example, Tencent's thesis is stored at reports/腾讯/腾讯-thesis.md. These files function as append-only logs, with each quarterly check adding new validation tables while preserving historical assumption states and health score trends.

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