// the find
xbtlin/ai-berkshire
AI 时代的伯克希尔：基于 Claude Code / Codex 的价值投资研究框架。巴菲特·芒格·段永平·李录四大师方法论 + 多Agent并行研究。| AI-era Berkshire: a value investing research framework built for Claude Code / Codex. 4 masters' methodologies + multi-agent adversarial analysis.
A collection of 18 Claude Code slash commands (markdown prompt files) that implement structured value investing research workflows based on four investors: Buffett, Munger, Duan Yongping, and Li Lu. The core mechanic is forcing the LLM to output buy/no-buy conclusions with price ranges instead of hedged analysis, using multi-agent parallel execution to get conflicting perspectives. Primarily aimed at Chinese-speaking investors analyzing Chinese and US equities.
The adversarial multi-perspective design is genuinely useful — having four agents with structurally conflicting frameworks (Buffett: 'cheap at 6x PE', Li Lu: '10-year certainty required, pass') surfaces real investment tension rather than producing another balanced non-answer. The financial_rigor.py tool uses decimal.Decimal throughout and enforces multi-source cross-validation with a >1% variance alert, which is the right instinct for a domain where float rounding kills people. The forcing functions are well-designed: the mirror test (can you justify the buy in 5 sentences?), information richness ratings (A/B/C) to prevent false confidence, and mandatory layered position sizing by risk tolerance. Over 100 actual research reports in the repo confirm this is a working tool, not a demo.
The 69%/66% annual return claims in the README are misleading — there's no control group, the framework was presumably iterated based on outcomes, and a single brokerage screenshot proves nothing about attribution. The repo recommends --dangerously-skip-permissions to reduce authorization prompts, which gets one paragraph of warning for a flag that disables all tool approval checks. There's no real data pipeline: agents search the web and draw on LLM training data, which makes analysis of fast-moving sectors unreliable — systematic backtesting is listed as a future TODO, meaning you can't currently verify whether the structured output actually improves decisions. The skills and all 100+ reports are in Chinese; the English README gives you the shape of the thing but not the actual prompts, so non-Chinese readers can't audit or extend the core workflows.