Portfolio Allocation and Rebalancing with the Portfolio-Review Skill in AI Berkshire

The /portfolio-review skill automates portfolio allocation and rebalancing by parsing your holdings, enriching them with live market data, and generating a six-step analysis that includes health checks, stress-testing, and cross-validated rebalancing recommendations approved by four value-investment masters.

The ai-berkshire repository is a modular research framework that implements value-investment principles—drawing from Warren Buffett (巴菲特), Charlie Munger (芒格), Duan Yongping (段永平), and Li Lu (李录)—through specialized AI agents. Its portfolio allocation and rebalancing with portfolio-review skill capability transforms single-stock research into comprehensive portfolio management via a declarative command that orchestrates data pipelines, scenario analysis, and master-level challenge protocols.

Architecture of the Portfolio-Review Skill

Three-Layer Design

The skill is defined in skills/portfolio-review.md and follows a strict three-layer architecture:

• Skill Layer – A declarative command (/portfolio-review) that parses user inputs (percentages, share counts, cash, or saved files) and triggers the analysis pipeline.
• Agent Layer – Background agents launched via the internal Task tool fetch live market data, valuation metrics, recent news, and analyst expectations for each holding in parallel.
• Tool Layer – Python utilities in tools/financial_rigor.py perform exact decimal calculations, cross-validate data sources, and run three-scenario valuation projections.

This separation ensures financial rigor while maintaining the flexibility to challenge recommendations from multiple investment perspectives.

Six-Step Analysis Pipeline

The pipeline defined in skills/portfolio-review.md progresses through six discrete stages:

1️⃣ Parse Holdings Inline parsing logic normalizes user input into a standardized table containing ticker symbols, share quantities, cost basis, current market price, percentage weight, and profit/loss calculations.

2️⃣ Data Enrichment Parallel Task agents retrieve current prices, PE/PB ratios, recent earnings, and key corporate events. The tools/financial_rigor.py module runs verify-valuation checks to ensure data integrity before analysis begins.

3️⃣ Single-Holding Health Check For each ticker, the skill asks three binary challenge questions (buy again? hold for five years? thesis still valid?) and produces a health-score table. This logic appears in lines 56-60 of the skill file.

4️⃣ Portfolio-Level Analysis The system constructs concentration matrices, correlation analyses, cash allocation summaries, and opportunity-cost comparisons. Key constraints—such as ensuring the largest position remains under 40%—are enforced here (lines 65-78).

5️⃣ Stress-Testing and Optimization Scenario analysis runs against global recession, US-China conflict, and interest rate spike conditions. The three-scenario valuation method from tools/financial_rigor.py ranks holdings by expected annualized return (lines 113-124 and 128-136).

6️⃣ Output Report Generation A structured markdown report is written to reports/portfolio-latest.md containing a concise health verdict, specific rebalancing actions (adds, cuts, cash moves), and a next-review reminder (lines 52-73).

Core Source Files and Components

Understanding the file structure is essential for customizing or debugging the allocation pipeline:

• skills/portfolio-review.md – The canonical skill specification defining input formats, the six-step pipeline, and output templates.
• tools/financial_rigor.py – Financial-validation utilities including verify-valuation for cross-source checks and three-scenario for probabilistic return projections.
• reports/portfolio-latest.md – Default output location for the generated portfolio analysis (created on each run).
• scripts/sync-codex-skills.py – Maintenance script that keeps the Codex skill artifacts in codex-skills/portfolio-review/SKILL.md synchronized with the markdown source.
• codex-skills/portfolio-review/SKILL.md – Generated Codex package derived from the source skill for IDE integration.

Running Portfolio Allocation and Rebalancing

CLI Usage with Claude Code

Install the Claude Code CLI and invoke the skill directly with your holdings:

# Install Claude client (see README for details)

npm install -g @anthropic-ai/claude-code

# Run the portfolio-review skill with a simple percentage list

claude /portfolio-review 腾讯30% 美团20% 茅台20% 英伟达15% 现金15%

CLI Usage with Codex

If you have installed the skill set via the provided installation script:

# After installing the Codex skill set

codex /portfolio-review "腾讯 500股 @480HKD, 美团 800股 @130HKD, 现金 20%"

Programmatic Invocation

You can trigger the skill from Python scripts using subprocess calls:

import subprocess, json

cmd = [
    "codex",
    "/portfolio-review",
    "腾讯30%, 美团20%, 茅台20%, 英伟达15%, 现金15%"
]
result = subprocess.run(cmd, capture_output=True, text=True)
print(result.stdout)   # displays the generated portfolio-review report

Summary

• The /portfolio-review skill automates portfolio allocation and rebalancing through a rigorous six-step pipeline defined in skills/portfolio-review.md.
• tools/financial_rigor.py provides the computational backbone with exact decimal arithmetic, verify-valuation cross-checks, and three-scenario probabilistic modeling.
• Agent Layer parallelism ensures real-time data enrichment without blocking the analysis workflow.
• Output reports in reports/portfolio-latest.md deliver actionable rebalancing instructions while preserving the four-master challenge philosophy—every recommendation is stress-tested against Buffett, Munger, Duan, and Li Lu's investment criteria.

Frequently Asked Questions

What input formats does the portfolio-review skill accept?

The skill accepts holdings expressed as percentages (腾讯30%), share counts with cost basis (腾讯 500股 @480HKD), absolute cash allocations (现金15%), or paths to saved portfolio files. The parsing logic in skills/portfolio-review.md automatically normalizes these formats into a standardized internal table.

How does the skill calculate rebalancing recommendations?

Rebalancing decisions combine concentration limits (e.g., ensuring a single position remains below 40% of total portfolio weight), opportunity-cost matrices, and three-scenario valuation rankings. The tools/financial_rigor.py module calculates expected annualized returns across bear, base, and bull cases, then suggests adds, trims, or exits based on risk-adjusted expected value.

Where is the portfolio analysis report saved?

By default, the skill generates a structured markdown report at reports/portfolio-latest.md. This file contains the health verdict for each holding, portfolio-level concentration analysis, stress-test results, and specific rebalancing actions with execution rationale.

Can I integrate the portfolio-review logic into my own Python codebase?

Yes. While the skill is designed for CLI invocation via Claude Code or Codex, you can import functions directly from tools/financial_rigor.py for custom workflows. The three-scenario and verify-valuation utilities are pure Python and can operate independently of the agent layer, though you will need to implement your own data fetching logic if bypassing the Task agents.

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