Content
English | [English](README_EN.md) | [日本語](README_JA.md)
[](https://trendshift.io/repositories/63696)
# Tool List
> "Price is what you pay, value is what you get." — Warren Buffett
>
> Redefining the depth and efficiency of investment research with AI.
**AI Berkshire** is a collection of investment research skills compatible with both Claude Code and Codex, systematically and structurally integrating the methodologies of four value investing masters: Buffett, Munger, Duan Yongping, and Li Lu. Achieve professional-grade investment research with AI Agents.
One person + Claude Code / Codex = A research team.
> 📮 **The repository contains the full framework, while the public account provides精选 research.** For companies truly worth in-depth research, along with my own judgments and choices beyond the reports, please follow the WeChat public account **复利炼丹炉** — [Scan the QR code to follow ↓](#精选研究首发于公众号)
[Real Track Record](#real-track-record) · [Why not directly ask AI](#为什么不能直接问-ai) · [Skills一览](#skills-一览20个) · [Quick Start](#快速开始) · [Real Research Reports](#实战研究报告) · [Design Philosophy](#设计理念) · [Public Account](#精选研究首发于公众号)
---
## Real Track Record
> Not just theoretical discussions. This framework is backed by a real investment system verified with actual money.
### 2024 Full-Year Returns: +69.29%
<img src="assets/2024-returns.jpg" width="300" />
### 2025 Full-Year Returns: +66.38%
<img src="assets/2025-returns.jpg" width="300" />
### Comparison with Major Indices
| Indicator | 2024 Full-Year | 2025 Full-Year |
|------|----------|----------|
| **This Framework** | **+69.29%** | **+66.38%** |
| Hang Seng Index | +17.67% | +27.77% |
| S&P 500 | +23.31% | +16.39% |
| CSI 300 | +14.68% | +17.66% |
| NASDAQ | +28.64% | +20.36% |
**2024 Excess Returns**: Outperformed S&P 500 by **46 percentage points**, outperformed Hang Seng Index by **52 percentage points**
**2025 Excess Returns**: Outperformed S&P 500 by **50 percentage points**, outperformed Hang Seng Index by **39 percentage points**
**Cumulative real track record over 2 years exceeds 1.46 million**, consistently outperforming major global indices for two consecutive years.
> *Disclaimer: Historical returns do not represent future performance. Screenshots come from a real Futu Securities account.*
### 精选研究首发于公众号
The repository contains the complete framework and full reports, while the public account provides **精选** — truly worth in-depth research companies, along with my own judgments and choices beyond the reports:
<img src="assets/wechat-qr.png" width="160" alt="微信公众号：复利炼丹炉" />
**复利炼丹炉** — Using AI to refine investment research.
---
## 为什么不能直接问 AI？
You can directly ask Claude: "Help me analyze if Pinduoduo is worth buying." You will get a balanced analysis that concludes with "Investment involves risk, please make your own judgment."
**This type of analysis looks correct but is not usable for decision-making.**
AI Berkshire solves not the "can it analyze" problem but the **analysis quality and decision-making discipline** issue. Here are the core differences:
### 1. Mandatory conclusions, no equivocation
Directly asking AI gives you a balanced "analysis." AI Berkshire outputs **pass / not pass / gray area**, with specific price ranges and tiered recommendations.
> Ordinary AI response: *"Pinduoduo has growth potential but also faces competitive pressure. Investors need to weigh..."*
>
> AI Berkshire Output:
> | Strategy | Recommendation | Price Range |
> |------|------|---------|
> | Aggressive | Build a 20% position at current price | $95-105 |
> | Cautious | Wait for policy clarification before building a position | $85-95 |
> | Conservative | Does not meet 10-year certainty criteria, observe | — |
>
> **Mirror Test**: If a 5-sentence answer is incomplete = do not buy, no exceptions.
### 2. Four Masters' perspectives vs. single analysis
Not just "analyzing with Buffett's method." Four perspectives create **real contradictions and tensions** —
For example, Pinduoduo:
- **Duan Yongping** (Business Model): Good business, C2M model hard to replicate → Score 3.7/5
- **Warren Buffett** (Financial Valuation): Low cash PE of only 6.3x, a cash machine → Score 4.4/5
- **Charlie Munger** (Contrarian Thinking): Moat shallower than imagined, Douyin reached 4 trillion GMV in 3 years → Score 3.5/5
- **Li Lu** (Long-term Certainty): Management culture has risks, 10-year uncertainty → Score 2.0/5
**Buffett says "really cheap," Li Lu says "uncertainty, don't buy"** — Such conflicts represent the true state of investment decisions. Single prompts cannot create such multi-perspective opposition, which is key to avoiding blind spots.
### 3. Structured anti-bias mechanism
The most dangerous thing about AI is not giving wrong answers but giving **answers that seem correct but do not stand up to scrutiny**. AI Berkshire has built-in multi-layer "anti-fraud" mechanisms:
| Mechanism | Solves what problem | Example |
|------|------------|------|
| **Information richness rating (A/B/C)** | Prevents the illusion that "more data = higher certainty" | Pop Mart rated B: limited data, estimated indicators with confidence intervals |
| **Munger-style reverse inspection** | Forces thinking of failure scenarios | "What scenarios would lead to Pinduoduo's demise?" → List 5 major scenarios and probabilities |
| **Quick rejection checklist** | 8 red flags for immediate rejection | Management integrity issues → Direct rejection, regardless of valuation |
| **Anti-consensus check** | Avoids aligning with market views | "What are smart people doing short?" → Discover overlooked risks |
| **Blank principle** | Prefer saying "don't know" | Data insufficiency marked as "gray area," no speculative certainty |
### 4. Financial data precision
LLM mental math is unreliable. A single decimal point error in PE or a mixed-up unit (e.g., HKD vs. CNY) can lead to incorrect investment decisions.
**Real case**: Analyzing Tencent, different sources provided market capitalization data in "billion HKD" and "billion CNY." AI Berkshire's approach:
```bash
# Manual market capitalization verification: share price × total shares, compared with report data
python3 tools/financial_rigor.py verify-market-cap \
--price 510 --shares 9.11e9 --reported 4.65e12 --currency HKD
# ✅ Verification passed, deviation only 0.08%
```
All calculations use Python `decimal.Decimal` (precise decimal), not `float`. Key data verified from at least two independent sources.
### 5. Reproducible research process
Directly asking AI, each output's format, depth, and coverage vary — Today analyzing Tencent includes moat scoring, but tomorrow analyzing Meituan might omit it.
AI Berkshire ensures: **same input → consistent output in structure and depth**. This means you can:
- Compare 7 companies horizontally with consistent scoring standards
- Re-analyze the same company after half a year and directly compare changes
- Align research results among team members
> Real output — 7 companies screened with the same Checklist:
>
> | Company | Pass? | Capability Circle | Good Business | Moat | Management | Margin of Safety | Comprehensive |
> |------|:-----:|:------:|:------:|:------:|:------:|:-------:|:----:|
> | Moutai | ✅ Pass | ★★★★★ | ★★★★★ | ★★★★★ | ★★★☆☆ | ★★★★☆ | 4.7 |
> | Tencent | ✅ Pass | ★★★★☆ | ★★★★★ | ★★★★★ | ★★★★★ | ★★★★☆ | 4.7 |
> | NVIDIA | ✅ Conditional | ★★★★☆ | ★★★★★ | ★★★★★ | ★★★★★ | ★★★☆☆ | 4.3 |
> | Meituan | ✅ Conditional | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★★☆ | 4.0 |
> | Kuaishou | ✅ Conditional | ★★★☆☆ | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★★★ | 4.0 |
> | Pinduoduo | ❓ Gray | ★★★★☆ | ★★★★☆ | ★★★☆☆ | ★★★☆☆ | ★★★★★ | 3.8 |
> | Pop Mart | ❓ Gray | ★★★☆☆ | ★★★★☆ | ★★★★☆ | ★★★★★ | ★★★☆☆ | 3.7 |
### 6. Multi-Agent parallelism = Research depth multiplication
`/investment-team` launches 4 independent Agents **simultaneously** researching a company. Each Agent searches, cross-verifies data, and gives independent conclusions. This is not splitting one prompt into four sections — 4 "analysts" each conduct complete research, and the Team Lead synthesizes.
One person directly asking AI has one context window. 4 Agents in parallel equal 4 times the search volume, 4 times the information sources, and 4 independent perspectives.
<p align="center">
<img src="assets/team-core.svg" alt="Team Lead scheduling four master Agents in parallel" width="720" />
</p>
### One-sentence summary
> **Ordinary people asking AI get "看起来对的分析" (看起来正确的分析), while using AI Berkshire gets "可以拿来做决策的投研报告" (usable research reports for decision-making).**
---
## Overall Architecture
<p align="center">
<img src="assets/architecture.svg" alt="AI Berkshire Overall Architecture" width="760" />
</p>
**Three-layer design philosophy**:
- **Skill Layer**: Abstract "what you want to do" into 20 clear entries — in-depth research, financial analysis, industry screening, portfolio management, thought tools, selected by scenario.
- **Agent Layer**: Team-type skills (e.g., `/investment-team`, `/earnings-team`) are scheduled in parallel by Team Lead with 4 master Agents — each searches, judges, and challenges independently, then synthesizes; lightweight skills directly connect to tools for quick in-and-out.
- **Tool Layer**: Precise calculations, real-time retrieval, report inspections — ensure data rigor and verifiability for each report.
---
## Skills一览（20个）
### 🔬 In-depth Research
| Skill | Purpose | Suitable Scenarios |
|-------|------|---------|
| [`/investment-research`](skills/investment-research.md) | Comprehensive in-depth analysis by four masters | Conduct comprehensive investment research on a listed company |
| [`/investment-team`](skills/investment-team.md) | Multi-Agent parallel research team | 4 Agents research in parallel, fastest and most comprehensive |
| [`/management-deep-dive`](skills/management-deep-dive.md) | In-depth management research | "Buying stocks is buying people" — deep dive when management is a core variable |
| [`/private-company-research`](skills/private-company-research.md) | In-depth research on unlisted companies | Research private companies like Ant Group and SpaceX with scarce information |
| [`/deep-company-series`](skills/deep-company-series.md) | 8-article series dissecting a company | Public account-level in-depth series, 120,000 words from cognitive reset to decision-making loop |
### 📊 Financial Analysis
| Skill | Purpose | Suitable Scenarios |
|-------|------|---------|
| [`/earnings-review`](skills/earnings-review.md) | Financial report reading (primary materials) | Read original financial reports, not dependent on second-hand research reports, like Buffett reading annual reports |
| [`/earnings-team`](skills/earnings-team.md) | Financial report reading team + public account release | Four masters interpret financial reports → editing and polishing → reader review → publishable articles |
### 🏭 Industry Screening
| Skill | Purpose | Suitable Scenarios |
|-------|------|---------|
| [`/industry-research`](skills/industry-research.md) | Industry chain panorama scan | Research all investment opportunities in an industry (slice by industry chain links) |
| [`/industry-funnel`](skills/industry-funnel.md) | Industry funnel screening | Entire market → coarse screening ≤10 companies → final selection 3 companies for in-depth analysis |
| [`/quality-screen`](skills/quality-screen.md) | Defect screening (7 hard indicators) | Quickly exclude non-first-class companies, support batch screening for stocks/industries/indexes/themes |
| [`/bottleneck-hunter`](skills/bottleneck-hunter.md) | Supply chain bottleneck hunter | From super trends, find physical bottlenecks and arbitrage opportunities in the supply chain |
| [`/investment-checklist`](skills/investment-checklist.md) | Buffett's pre-buy checklist | 6 gates for quick screening, 10 minutes to decide if worth in-depth research |
### 📈 Portfolio Management
| Skill | Purpose | Suitable Scenarios |
|-------|------|---------|
| [`/income-investment`](skills/income-investment.md) | Yield stock analysis | Distinguish sustainable income, opportunistic high dividend, and yield traps |
| [`/portfolio-review`](skills/portfolio-review.md) | Portfolio management and optimization | From "research companies" to "manage portfolios" — positions, concentration, rebalancing |
| [`/thesis-tracker`](skills/thesis-tracker.md) | Investment thesis tracking | Post-buying discipline system: continuously track if the thesis is disproven |
| [`/thesis-drift`](skills/thesis-drift.md) | Investment thesis drift detection | Compare two theses/reports, distinguish fact changes, valuation changes, and wording changes |
| [`/news-pulse`](skills/news-pulse.md) | Stock price abnormality quick attribution | Within 10 minutes, clarify "what happened" during large price increases/decreases |
### 🧠 Thought Tools
| Skill | Purpose | Suitable Scenarios |
|-------|------|---------|
| [`/dyp-ask`](skills/dyp-ask.md) | Duan Yongping Q&A | Think with Duan Yongping — business, investment, life |
| [`/financial-data`](skills/financial-data.md) | Financial data retrieval and cross-verification specifications | Ensure key data from 2 independent sources, error >1% alert |
| [`/wechat-article`](skills/wechat-article.md) | WeChat public account articles | Authors, editors, and readers collaborate to produce publishable articles |
---
## Quick Start
### Cost and Model Selection
In-depth research skills default to multi-round research, cross-verification, and multi-Agent comprehensive judgment, so token consumption is higher, which is to obtain more complete business, financial, industry, and risk analysis.
For actual investment decisions with high risk and importance, the view of the maintainer is that the strongest model usually brings better analysis ROI, and it is not recommended to sacrifice key judgment quality for model cost savings. Lightweight models are more suitable for initial screening, summaries, or low-risk issues; when involving moat, valuation, management, and risk cross-judgment, expect analysis quality to depend more on model capabilities.
When controlling costs, prioritize workflow adjustments rather than expecting complete in-depth research to become cheaper: quickly exclude companies with [`/quality-screen`](skills/quality-screen.md), and use [`/news-pulse`](skills/news-pulse.md) for price anomaly attribution. Only when the result is worth further research, then run [`/investment-research`](skills/investment-research.md) or [`/investment-team`](skills/investment-team.md).
### 1. Install AI Client
This repository retains the same canonical workflow and provides Claude Code commands and Codex skills. Install according to your client.
Claude Code users:
```bash
npm install -g @anthropic-ai/claude-code
```
Codex users:
```bash
# macOS / Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# or use npm
npm install -g @openai/codex
# or use Homebrew
brew install --cask codex
# Verify installation
codex --version
```
Windows users can use the official PowerShell installation command: `powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"`.
If `codex --version` outputs the version number, you can continue installing Codex skills.
#### Reduce authorization confirmation
These skills frequently call tools, and Claude Code defaults to request authorization confirmation each time. This behavior comes from Claude Code's permission mechanism, which is not modifiable by this repository.
If you trust the current workflow and run it in a trusted environment, you can start Claude Code with the permission confirmation skipping mode:
```bash
claude --dangerously-skip-permissions
```
Note: This mode disables Claude Code's tool approval protection and should only be used when you trust the repository, commands, and working directory.
### 2. Install Skills
Claude Code users install (macOS / Linux):
```bash
# Clone repository
git clone https://github.com/xbtlin/ai-berkshire.git
# Copy skills to Claude Code global commands directory
cd ai-berkshire
./scripts/install-claude-commands.sh
```
Claude Code users install (Windows PowerShell / Command Prompt):
```bat
git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire
.\scripts\install-claude-commands.bat
```
Codex users install (macOS / Linux):
```bash
# Clone repository
git clone https://github.com/xbtlin/ai-berkshire.git
# Generate and install Codex skills to ~/.codex/skills
cd ai-berkshire
./scripts/install-codex-skills.sh
# Optional: Install Codex slash prompts to ~/.codex/prompts
```
Tool List
## For a /investment-research experience close to Claude Code
./scripts/install-codex-prompts.sh
Codex user installation (Windows PowerShell / Command Prompt):
```bat
git clone https://github.com/xbtlin/ai-berkshire.git
cd ai-berkshire
.\scripts\install-codex-skills.bat
REM Install Codex
.\scripts\install-prompts.bat
repository maintains three sets of entries: `skills/*.md the source file of command; `codex/*/SKILL.md` Codex skill package `scripts/sync-skills.py` from `skills/*.md`; `codex-prompts/*.md` is an optional Codex slash prompt compatibility layer.
### 3. Usage
Call directly in:
```bash
# In-depth research
/in-research Tencent
/in-team Meituananagement-deep-diveituan
/private-company-research SpaceX
/deep-company-series Pinduoduo
# Financial report analysis
/earnings-review Tencent 2025Q4
/earnings-team PDD 2025 annual report
# Industry screening
/industry-research nuclear power
/industry-funnel AI computing power
/quality-screen Hang Seng Index constituents
/bottleneck-hunter AI infrastructure
/investment Moutai, NVIDIA, Apple
# Position management
/income Verizon mode=existing role=core-income quantity cost_basis=39 tax_residence=France horizon=5y
/portfolio 30%, Me 20%, M 20%, cash 30%
/thesis-tracker Pinduoduo
/rift Pinduoduo reports/Pinduoduo-th2025Q4.md reports/Pinduoduo-th2026Q1/news-pulse Tencent
tools
/dyp What is Pinduod moat?
/-article Meituan
```
After installing Codex, restart Codex, and then task directly by skill for example:
```text investment-research to
Use earnings-review to analyze PDD's 2025 annual report
Use industry-funnel to screen AI computingUse bottleneck-hunter AI infrastructure bottlenecks thesis-drift to compare two investment papers on PinduodUse wechat-article to write a Meituan research article
```
If slash prompts are installed, you can also search for these prompts in the `/` menu after restarting Codex. The official Codex custom prompt entry usually displays as `prompts>`, for example:
```text
/prompts:earch Tencent
```
## Detailed Introduction to Each Skill 1. `/investment` — Comprehensive Analysis Masters
The most comprehensive single-company in-depth research Execute in seven modules in order:
```
Data collection → (Duan Yongping) → Moat () → Reverse thinking (Munger)
→ Management assessment (Duan Yongping + Buffett) → Civilization trend (Li Lu) → Valuation and safety margin
```
****:
- AI research bias self-awareness mechanism (/C level information richness rating)
- Key data multi-validation (market capitalization manual verification, at least 2)
- Four mastersquestioning" throughout the article
- Three-soptimistic / neutral / pessimistic) + reverse DCF
**Output example excerpt**:
> #### Comprehensive decision memorandum
>
> | Dimension | Conclusion | Confidence |
> |------|------|--------|
> | Business quality (Duan Yongping) | Excellent: platform-type business, bilateral network effects, marginal cost tends to zero★★★★ |
> | Moat (Buffett) | Wide and network effects + conversion costs + scale effects superimposed | ★★★ |
> | Managementuan Yongping + Buffett Excellent: founder helm, strong capital allocation discipline |★☆ |
> risk (Munger Regulatory policy uncertainty, new business losses dragging down overall ★★★☆☆ |
> | Civilization trend (Li Lu) | Conform to digital consumption trend, but notization-level paradigm shift ★★★★☆ | Valuation (Buff Duan Yongping Current PE 18 in historical median low, safety margin | ★☆ |
>
>uan Yongping**: essence of this business connect consumers and merchants, making efficiency improvement money. A good business is marked by: the more users, the more merchants; the more merchants, users. Once the flywheel turns, it's hard to stop."
>
Munger**: "Think backwards — if disappears tomorrow, users and merchants do the answer is 'find alternatives quickly', then the is not deep enough. If the answer is 'life will become very inconvenient', then it's worth paying attention."
---
### 2. `/investment-team` — Multi-Agent Research Team
Launch 4 AI Agents to study in parallel, simulating a real investment research team collaboration. Each Agent searches, analyzes, and scores independently, and then the Team Lead makes a comprehensive judgment.
**Output example excerpt**:
> #### conclusion
> Meituan is China's absolute leader in local life services, with multiple network effects moat, current valuation at historical low, significant long-term investment value, recommended to buy on dips.
>
> #### Four-dimensional scoring table
>
> | Dimension | Framework | Score | Core judgment |
> |------|------|------|----------|
> | Business model & Moat | Yongping | ★★★★☆ | Strong bilateral network effects, forming a flywheel |
> | Finance & Valuation ★★★★☆ | Core business profit margin continues to improve, valuation at historical low |
> | Industry & Competition | Munger | ★★★☆☆ | Douyin invades into-store business, risk of deterioration in competition pattern |
> | Risk & Management | Li Lu | ★★★★☆ | Wang Xing's strategic vision is excellent, but new business burning requires vigilance |
**Comprehensive score: 3.8 / 5**
>
> #### Investment advice
>
> | Strategy | Advice | Price range (Hong Kong dollars) |
> |------|------|---------------|
> | Aggressive | Build a 30% position at current price | 120-140 |
> | Conservative | Wait for a callback to 100-110 to build a position | 100-120 |
> | Conservative | Wait for quarterly report to verify profit margin trend before intervening | <100 |
---
### 3. `/investment-checklist` — Buffett's Pre-Buy Checklist
A six-step quick screening to help you decide whether a company is-depth research within 10 minutes:
```
First gate circle (can I understand it?)
↓ Pass gate: good business (what are the economic characteristics?)
↓ Pass
Third gate: moat ( is the competitive advantage?)
↓ Pass
Fourth gate: management (is it trustworthy?)
↓ Pass
Fifth gate: safety margin (is the price cheap?)
↓ Pass
Sixth gate: decision-making discipline (is it rational or FOMO?)
↓ Pass
✅ Mirror test
```
**Supports multi-company comparison** — screen multiple targets at once:
```
/investment-checklist Tencent, Alibaba, Meituan, Pinduoduo
```
**Output example excerpt**:
> #### Mirror test
>
I bought Tencent at 380 Hong Kong dollars because 1. The essence of this business is **social network + digital content platform**, I understand it;
> 2. Its moat is **12 billion users' social relationship chain**, and it's widening;
> The management team **Pony Ma is low-key and pragmatic, with excellent capital allocation**, trustworthy;
> 4. The current price is equivalent to **80% of intrinsic value**, with a certain safety margin;
> Even if I'm wrong, the downside risk is controllable account has a net cash of over , and the game cash flow is strong**."
>
> ✅ Pass the mirror test
>
> **If you can't complete the 5 sentences = don't buy. No exceptions.**
---
### 4. `/industry-research` — Industry Chain Panoramic Scanning
Starting from an investment theme, complete an industry chain panoramic research:
```
Investment logic chain construction → Industry chain panoramic diagram → Global listed company scanning
→ Head companies in each link four-master analysis → configuration suggestions
```
**Output example excerpt**:
> #### Investment logic chain: nuclear power
>
> Underlying trend: AI data center power demand surge + carbon neutrality target
> → Leads to: stable clean base load power demand surge
> → Creates: nuclear power new construction / SMR's deterministic demand
> → Benefits: uranium mining → fuel processing → equipment manufacturing → operator
>
> #### Recommended portfolio
>
> | Level | Position | Target | Core logic |
------|------|------|------|---------|
> | Core |% | China Nuclear Power (CGN), Cameco | Operation + uranium mining | Highest certainty |
> | Satellite | 30% | China Nuclear Power, Dongfang Electric | Operation + equipment | Domestic substitution benefits |
> | Option | 15% | NuScale, Nano Nuclear | SMR | High risk high elasticity |
> | ETF | Alternative | URA, URNM | Full chain | Lazy solution |
---
### 5. `/industry-funnel` — Industry Funnel Screening
Starting from an industry / direction, **global market → ≤10 companies → 3 companies** layer by layer selected:
```
Global market scanning (activity + increase + market value top , 30-60 companies)
↓ Value5 hard indicators
Coarse screening ≤ 10 companies
↓ Fine analysis (each 300-500 words)
Fine analysis ≤ 10 companies
↓ Final selection (according to portfolio complementarity, not scoring top 3)
Four-master in-depth analysis 3each 800-1200 words)
↓
Recommended portfolio (core / satellite / option) + operation signal
```
**Core features**:
- Each layer retention / abandonment and eliminated targets leave (not black box)
- Final selection3 companies according to "portfolio complementarity" ( + medium elasticity + high elasticity), not sorted top 3
- Forced listing of IPO candidates" to avoid missing core players in the primary market
- AI bias self-awareness mechanism: respond to leader preference / English preference / story preference / listing preference
**Difference from `/industry-research`**:
- `industry-research` focuses on industry chain structure and panorama (sliced by link)
industry-f focuses on individual stock screening funnel (layer by from the global market to 3 companies)
test: AI industry 4 sub-track parallel6-05-09)**:
| Sub-track | Final selection 3 companies | Core position recommendation |
---------|------------|
| AI computing power | TSMC / SK Hyn TSMC ★★★★★ AI model | Alphabet / Meta / Alibaba |★★★★ |
| AI application | Microsoft / AppLovin | Microsoft + Adobe ★★★★ AI infrastructure power | Eaton / TBE / Energy | EatonBE ★★★★Key findings**: The biggest winner of AI application not AI Native companies with channel + workflow embedding degree — this echoes the 1990 internet bubble "selling shovels law (Amazon and Apple win, Pets.comComplete report: [AI computing power](reports/A20computing%20power-funnel-202) · model](reports/Amodel-funnel-20260509.mdAI application](reports/AI%20unnel-20260509.md) · [AI](reports/A20infrastructure%20power-funnel-20260509.md)
---
### 6. `/private-company-research` — Unlisted Company In-Depth Research
A "detective-style" research framework designed for unlisted companies with scarce information:
**Core differentiation**:
- **Financial data piecing together**: from prospectus, parent company financial report, financing news, industry data multi-source piecing
- **Confidence level marking**: each data point marked with 🟢high / 🟡medium / 🔴low confidence level
- **Multi-method valuation cross-validation**: financing valuation method + comparable company method + DCF + terminal deduction method
- **Exit path analysis**: IPO / M&A / secondary transfer full path assessment
**Output example excerpt**:
> #### Company profile snapshot: SpaceX
>
> | Project | Content |
> |------|------|
> | Latest valuation | ~$350B (2025 secondary market) 🟡 |
> | Estimated revenue | ~$13 billion (2024) 🟡 |
> | Starlink users | 4 million+ (end of 2024) 🟢 |
> | Launch frequency | 100+ times/year (2024) 🟢 |
>
> #### Valuation judgment
>
> | Method | Valuation range | Explanation |
> |------|---------|------|
> | Recent financing | $350B | Secondary market quotation, with liquidity premium |
> | Comparable company method | $200-280B | Benchmarking telecom + aerospace + defense |
> | DCF (neutral) | $250-350B | Assuming Starlink $300 billion revenue in 2027 |
> | Terminal deduction | $400-600B | Assuming Starlink becomes global telecom infrastructure |
>
> **Comprehensive reasonable valuation range: $250B - $400B**
---
### 7. `/news-pulse` Price Abnormal News Attribution
A situational response skill designed for quickly figuring out what happened when stock prices surge or plummet. **Not in-depth research, but 10-15 minute quick attribution** — avoid being anxious about small essays or blindly selling.
**Core differentiation**:
- **4-dimensional parallel reconnaissance**: company event / regulatory policy / industry opponent / market sentiment (sell-side + big V + southbound funds)
- **Attribution prioritizes over listing**: not listing all news, but judging "which event deserves this stock price movement"
- **Forced nature judgment**: value event / emotional fluctuation / **real cause unknown** / mixed — where "real cause unknown" is the most valuable output (may have insider rushing)
- **Clear action suggestion**: whether to trigger in-depth research, whether to revisit the thesis, whether to observe only, etc.
**Difference from other Skills**:
| Scenario | What to use |
||
| Complete research (hour level) | `/investment-team` or `/investment-research` |
| Financial report in-depth reading | `/earnings-review` |
| Long-term thesis tracking | `/thesis-tracker` |
| **10-minute stock price abnormal attribution** | **`/news-pulse`** |
**Output example excerpt** (Tencent 4/17-5/01 actual test, 14 days -10.47%):
> #### One-sentence attribution
> This -10.47% drop is about 70-80% driven by fund flow + sentiment, and 20-30% by deferred digestion of AI investment doubling — **no fundamental negative**, sell-side maintains buy consensus, and the nature is a "liquidity + emotional callback", not a value event.
> #### Abnormal attribution table
>
> | Candidate explanation | Estimated contribution | Confidence |
> |---------|--------|--------|
> | Disappearance of repurchase silence period (5/13 before financial report) | -3% ~ -4% | High |
> | Southbound funds turned to net sell Tencent | -2% ~ -3% | High |
> | AI narrative was taken away by competitors (DeepSeek V4/Qwen3.6/month dark 1T) | -1% ~ -2% | Medium |
> | Sector / macro beta (oil + geopolitics + Fed Warsh hawkish) | -2% ~ -3% | High |
> | Pre-quarter caution | -1% ~ -2% | Medium |
> | Fundamental deterioration | **0%** | Extremely high (excluded) |
>
> #### Nature judgment: ✅ Mixed
> 70% liquidity / sentiment + 20% long-term AI narrative concerns + 10% pre-quarter uncertainty
>
> **Key evidence against**: Duan Yongping sold Tencent put (looks bullish) on 4/ sell-side 24 firms' consensus; rose 2% on 4/30 (excluded game); Tencent outperformed Hang Seng 7 percentage points (Hang Seng actually rose this month).
Invocation:
```
/news-pulse Tencent
/news-pulse Pinduodmet 12% within one week
/news-pulse Meituan
```
## Practical Research Reports
> The following are real reports generated using this framework, actual output effects of AI research.
| Company | Used Skill | Core Conclusion | Report Link |
|------|-----------|---------|---------|
| Pinduoduo (PDD `/investment-team` | Comprehensive 3.4/5, extremely cheap10-year certainty insufficient, suitable for | [View Report](reports/) |
| Tencent Holdings (0700.HK) | `/investment-research` | Social monopoly + allocation, 14x forward PE reasonable and low, | [View Report](reports/) |
| 7 companies comparison | `/investment-checklist` | Moutai, Tencent pass; NVIDIA, Meituan, Kuaishou; Pinduoduo, Pop | [View Report](reports/20260408.md) |
| Master tracking | Custom research | Latest 13F holdings of Buffett / Li Lu / Duan Yongping + PDD cost analysis | [View Report](reports/20260408-research-20260408.md) |
> *More reports will be added. Welcome to submit your research reports generated using this framework.*
---
## Design Philosophy
### Integration of Four Masters' Methodologies
**Duan Yongping · ""** — The essence of business is the common starting point of the other three perspectives:
| Buffett | Munger | Li Lu |
|:---:|:---:|:---:|
| Moat<br>Safety margin<br>Management | Reverse thinking<br>Risk list<br>Bias self-examination | Civilization trend<br>Paradigm shift<br>Industry value |
The four masters are not simply divided, but designed to **challenge each other**:
- Duan Yongping says "", Munger will ask will it die"
- Buffett says "cheap enough", Li "will it still be 10 years"
- You four reports spliced together, but of four thinking modes
### Financial Tool (`tools/financial_rigor.py`)
| Function | Command | |
|------|------|-----------|
Market capitalization calculation** | `verify` | Accurate calculation of share total shares, detecting unit errors |
Valuation calculation** | `` | Accurate decimal calculation ofB/ROE/FCF| **Multi-source cross-validation `cross-validate` | Automatic comparison same data from N sources, alerting if exceeding tolerance |
| **Three valuation** | `three-s | Accurate calculation of target price under neutral / pessimistic assumptions |
Benford's Law detection**benford` | Detecting abnormal first digits in financial data |
|urate calculator** | `calc Accurate calculation of arbitrary financial expressions LLM mental math |
**Design principles**: All calculations use Python `decimal.Decimal` (accurate decimal), not `float` (floating-point approximation). `0.1 + 0.2 = 0.3` is not allowed to fail in financial scenarios.
---
## Future Directions
- [ ] Historical backtest: AI research report vs. actual stock price performance
- [ ] Macroeconomic cycle
- [ ] MCP-based data access (Wind/Bloomberg/Yahoo Finance)
---
## Disclaimer
is for learning and research purposes only does not constitute any investment advice. Investment carries risks, and decisions. Please always do your own due diligence (DYOR).
## License
MIT License
---
> "The best investment you can make is in yourself." — Warren Buffett
>
> AI Berkshire: Empowering everyone with their own research team.
## Star History
If this project has been helpful to you, please give it a Star! The refined company research and personal judgment were first published on the WeChat public account **Fu Li Lian Dan Lu** (QR code at [the beginning of the article](#精选研究首发于公众号)).
<a href="https://github.com/xbtlin/ai-berkshire/stargazers">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="assets/star-history-dark.svg">
<img alt="Star History Chart" src="assets/star-history.svg">
</picture>
</a>
