AI Berkshire
A structured value-investing research workflow for Claude Code and Codex.
Per-dimension scores and reasoning
Evidence shows: README explicitly instructs users to use --dangerously-skip-permissions mode, which weakens user confirmation, but the repo does not enforce it and warns of risks. Data flow transparency: README describes multi-agent parallel search and cross-validation, but does not detail data flow and storage. Sensitive data handling: No code handling sensitive data found, but tools may read financial data. Dependency security: No dependency manifest found, cannot assess. External effects: Tools call external APIs and web searches, but permission scope not specified. Rollback: No rollback mechanism provided. Source attribution: README claims real track record but provides no verifiable audit trail. Deductions: Lack of explicit least-privilege design, incomplete user confirmation, opaque data flow, no rollback, and unsubstantiated source attribution.
Evidence shows: README and test files demonstrate tool consistency, tests cover encoding and number extraction bug fixes. Dependency availability: No dependencies listed, but tests show only standard library usage. Failure messages: Tests verify error paths, but not all failure scenarios. Deductions: Missing dependency manifest, incomplete failure message coverage.
Evidence shows: README provides multiple usage scenarios and examples, covering different user needs. Capability boundaries: README describes each skill's purpose and limitations, but boundaries not explicit. Trigger precision: Command examples are clear, but precise trigger conditions not defined. Environment fit: Multi-platform installation instructions provided, but not all environments tested. Deductions: Capability boundaries and trigger precision not precise enough.
Evidence shows: README is well-structured, provides installation instructions, examples, and FAQ. Naming stability: Skill names are consistent. Known limitations: README mentions high token consumption, but not other limitations. License: MIT license clear. Versioning/changelog: No changelog provided. Maintenance responsibility: Maintainer not identified. Deductions: Missing changelog and maintainer information.
Evidence shows: README displays detailed output examples, practical. Marginal value: Multi-agent parallel and cross-validation add value. Cost-benefit: README mentions high token consumption, but no cost estimates. Deductions: Cost-benefit analysis insufficient.
Evidence shows: README claims real track record, but no verifiable data sources. Cross-validation: Tools support multi-source cross-validation, but actual validation results not shown. Fact/inference separation: README distinguishes facts and inferences, but not systematically. Deductions: Performance claims lack traceability, cross-validation results not demonstrated.
- Real track record claims lack verifiable audit evidence and should not be relied upon for reliability.
- Users should be cautious with --dangerously-skip-permissions mode and ensure trusted environment.
- Dependency manifest missing, supply chain security cannot be assessed.
What does this agent do, and when should you use it?
AI Berkshire is a collection of value-investing skills for Claude Code and Codex. Its 20 task-specific entry points cover company research, earnings review, industry screening, portfolio work, thesis tracking, and WeChat article production. Heavier workflows use a Team Lead to coordinate four independent master-inspired perspectives, while lighter skills run their own direct workflow. The repository maintains Claude Code command sources, generated Codex skill packages, and an optional Codex slash-prompt compatibility layer. Its outputs are research reports, screening results, valuation calculations, and suggested actions; the project states that it is not investment advice.
Users invoke workflows with commands such as /investment-research 腾讯, /investment-team 美团, or /earnings-review 腾讯 2025Q4. /investment-research organizes work into seven stages: data collection, business essence, moat analysis, inversion, management assessment, civilization trends, and valuation with margin of safety; /investment-team runs four independent perspectives in parallel and has a Team Lead synthesize them. /industry-funnel narrows a market scan to no more than 10 companies and then three final selections, while /portfolio-review, /thesis-tracker, and /news-pulse address post-investment management and price-move attribution. tools/financial_rigor.py exposes verify-market-cap, verify-valuation, cross-validate, three-scenario, benford, and calc, using Python decimal.Decimal for precise decimal calculations. Source workflows live in skills/*.md, and scripts/sync-codex-skills.py generates codex-skills/*/SKILL.md packages from them.
- An individual investor preparing a long-term study of a listed company can use
/investment-researchto produce a memo spanning business quality, moat, management, risks, and valuation. - A researcher who needs several independent perspectives quickly can run
/investment-teamfor parallel analysis of one company and review the Team Lead synthesis. - An investor reviewing an earnings release from primary materials can use
/earnings-reviewor/earnings-teamfor a named company and reporting period. - Someone investigating an investment theme such as AI compute or nuclear power can use
/industry-researchfor a value-chain view or/industry-funnelto select candidates from a market scan. - A holder facing a sharp price move can use
/news-pulseto assess company events, regulation, competitors, and market sentiment. - An investor managing several positions can use
/portfolio-review,/thesis-tracker, and/thesis-driftto monitor allocations and test whether an investment thesis has changed.
What are this agent's strengths and limitations?
- It separates company research, earnings analysis, industry funnels, portfolio management, and thesis-drift checks into 20 named skill entry points.
/investment-teamand/earnings-teamuse four parallel master-inspired perspectives plus Team Lead synthesis instead of a single-perspective workflow.tools/financial_rigor.pycovers market-cap checks, valuation checks, multi-source comparison, scenario valuation, Benford analysis, and precise calculations withdecimal.Decimal.- It offers Claude Code commands, generated Codex skills, and optional Codex slash prompts for the two explicitly supported clients.
- The README states that deep-research workflows consume substantial tokens because they involve repeated research, cross-validation, and multi-agent synthesis.
- Research quality is described as model-dependent, and the maintainer recommends stronger models for moat, valuation, management, and risk judgments.
- Core usage depends on Claude Code or Codex; the repository does not provide a general API, standalone web app, or deployment path for other runtimes.
- Real-time MCP data access remains a future direction; no ready-to-configure Wind, Bloomberg, or Yahoo Finance integration is provided.
- The project produces research and suggested price ranges, but explicitly disclaims investment-advice status and asks users to conduct their own due diligence.
How do you install or deploy this agent?
Install and verify Codex first:
npm install -g @openai/codex
codex --versionThen install the repository's Codex skills:
git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire
./scripts/install-codex-skills.shOptional slash prompts:
./scripts/install-codex-prompts.shFor Claude Code, run this after cloning:
./scripts/install-claude-commands.shThe README does not specify any dedicated API keys or financial-data credentials.
How do you use this agent?
Restart Codex after installation, then name the skill and target in a request, for example: 使用 investment-research 研究腾讯, 使用 earnings-review 分析 PDD 2025年报, or 使用 industry-funnel 筛选 AI算力. With Codex prompts installed, use /prompts:investment-research 腾讯. In Claude Code, enter commands such as /investment-team 美团 or /investment-checklist 腾讯, 阿里巴巴, 美团, 拼多多.
How does this agent compare with similar options?
Compared with directly asking an AI whether a stock is worth buying, AI Berkshire adds fixed skill workflows, information-richness ratings, rapid-veto checks, inversion, and multi-perspective synthesis to constrain the research process and output. It is not a standalone replacement for Claude Code or Codex; it supplies investing workflows within those clients.
FAQ
Is deep research expensive to run?
/quality-screen and /news-pulse are presented as lighter first steps.