xbtlin/ai-berkshire

Purpose and Entry Points

AI Berkshire is a repository of AI-agent investment-research workflows, reusable skills/prompts, precision financial tools, and generated research artifacts. It is not a deployed application or service. Users install skills into Claude Code or Codex, then invoke research workflows by name. Primary entry points are skills/*.md for Claude Code, codex-skills/*/SKILL.md for Codex, and optional Codex slash prompts in codex-prompts/*.md.

Canonical Workflow Distribution

The Claude-oriented Markdown skills are the canonical workflow definitions; Codex skill packages are generated copies. Installation scripts put these assets into client-specific global directories, while synchronization scripts maintain parity.

• skills/investment-research.md

• scripts/sync-codex-skills.py

• scripts/install-codex-skills.sh

This boundary matters: workflow edits should originate in skills/, then be propagated rather than independently changed in generated Codex skill directories.

Research Orchestration Layer

Skills encode task-specific control flows for company research, earnings review, industry discovery, screening, portfolio review, news triage, and publication. The deepest flows use four independent value-investing perspectives—business model, financial/valuation, contrarian risk, and long-term trend—then synthesize a decision. Team-oriented workflows delegate parallel agent work; lightweight workflows go directly to tools and structured outputs.

• skills/investment-team.md

• skills/earnings-team.md

• skills/industry-funnel.md

Data Acquisition and Quantitative Validation

The tool layer provides local Python CLIs for financial calculations, data retrieval, screens, backtests, and report checking. financial_rigor.py is the central correctness boundary: it uses decimal.Decimal for market-cap, valuation, cross-source, and scenario calculations, preventing LLM arithmetic or unit errors. Data inputs are persisted locally as CSV/JSON rather than behind a database.

• tools/financial_rigor.py

• tools/stock_screener.py

• data/fundamentals.json

External data is obtained by tools or agent research from market/public sources; visible integrations include Morningstar-oriented data, A-share/Taiwan-stock access, and Xueqiu scraping. No always-on data pipeline or credentials-managed service is represented in the tree.

• tools/morningstar_fair_value.py

• tools/ashare_data.py

• tools/xueqiu_scraper.py

Evidence, Audit, and Reproducibility

High-stakes reports commonly preserve a research package alongside the prose: original PDFs/text, source registries, assumptions, model inputs, calculation outputs, audit samples, verdicts, and Python reproduction scripts. This creates a file-based provenance boundary between claims, inputs, and independently rerunnable verification.

• reports/腾讯/审计-20260906/复核计算.py

• reports/泡泡玛特/审计-20260906/audit-results.json

• reports/快手/《看懂快手》-20260906/审计/reproduce.py

Generic report-level quality checks are implemented separately and tested, supporting consistency across the large report corpus.

• tools/report_audit.py

• tests/test_financial_rigor.py

• tests/test_report_audit.py

Research Corpus and Indexing

reports/ is the principal product store: Markdown investment reports organized by company, sector, screening universe, and recurring research series. It also contains generated audit evidence and occasional publishable PDF artifacts. Index files expose the corpus; reports_index.py updates README/index material rather than serving a runtime search API.

• reports/README.md

• reports/index.json

• tools/reports_index.py

Runtime and Operational Model

The material runtimes are Python 3 for tools, synchronization, audits, and report builds; shell/batch scripts for installation; and user-installed Claude Code or OpenAI Codex clients for agent execution. One specialized publication pipeline builds a bilingual survey manuscript into LaTeX/PDF outputs.

• reports/RSI-survey-20260906/build.py

• reports/RSI-survey-20260906/arxiv/main.tex

• output/pdf/rsi-survey-en-20260906.pdf

There is no web server, database, queue, container, or cloud infrastructure represented. GitHub provides repository hosting and issue intake; agent clients and external public/market-data sources are the primary execution dependencies.