AI Berkshire
Claude Code skill framework for value-investing research; useful to Kevin mainly as a pattern for domain-specific skill packs with adversarial specialist agents and deterministic calculation tools.
What it is
xbtlin/ai-berkshire packages investment research into a Claude Code skill suite. The README frames it as a value-investing research framework built around Buffett, Munger, Duan Yongping, and Li Lu methodologies, with multiple agents researching independently and a team lead synthesizing the output. Source: GitHub README, 2026-06-26
The reusable architecture is three-layered:
| Layer | Pattern |
|---|---|
| Skill layer | 16 explicit task entries for company research, earnings review, industry screening, portfolio review, thesis tracking, news attribution, and thinking tools |
| Agent layer | four specialist perspectives run in parallel, challenge each other, then synthesize |
| Tool layer | exact financial calculations, live retrieval, and report checks to reduce hallucinated numbers |
Source: GitHub README, 2026-06-26
Why it matters
Do not treat this as a default investment-advice system. Treat it as evidence for a stronger agent pattern:
- domain expertise lives in named skills, not one huge prompt
- specialist agents should disagree on purpose
- numeric domains need deterministic tools, not model mental math
- decisions need explicit
pass / fail / gray areaoutputs instead of balanced but actionless summaries - low-confidence data should be labeled at the datapoint level
That pattern ports well to Kevin's own agent-doc packs, product research, diligence, and eval loops.
Routing
Use AI Berkshire as a reference when designing a domain skill suite that needs:
- multiple expert lenses
- adversarial critique
- explicit recommendation formats
- calculations or data validation tools
- an audit trail from claim to source
For actual financial decisions, re-check all live market data, sources, and risks. Historical returns in a README are not a decision rule.
Timeline
- 2026-06-26 | Promoted from Kevin's GitHub stars harvest. GitHub API reported 3,052 stars, MIT license, Python, and topics including
claude-code,financial-analysis,investment-research,multi-agent,mcp, andvalue-investing. Source:raw/github/xbtlin_ai-berkshire-repo-2026-06-26.json; GitHub API, 2026-06-26