Organize all of fund manager Zheng Xi’s quarterly reports, notes, and interviews from the past decade into a structured corpus. Build it into a traceable AI skill so that AI can do research Q&A and fund analysis based on real corpus, not model hallucinations. https://t.co/Rkbl06soCU https://t.co/GF8SyI0JYK — # lyra81604/zhengxi-views Source: https://github.com/lyra81604/zhengxi-views # 🔎 Fund Manager · Zheng Xi Views Library Skill ### Traceable · Based on Original Texts · Connected to All-Market Funds · Cross-AI Platform “Check views · Learn methods · Make forward-looking judgments · Imitate tone for commentary · Cross-reference words and actions · Compare and score across all market funds — let any AI understand E Fund’s Zheng Xi”

Install · What It Can Do · Use in Other AI · Directory Structure · Data Sources · Limitations

Many answers about “What does fund manager X think of Y” sound plausible but are actually fabricated by the model from memory—there is no record of such statements. This skill solves that problem: its foundation is the original texts left by Zheng Xi (Deputy General Manager of E Fund Equity Investment Management Department, Fund Manager). Every conclusion can be traced back to “which year, which document, what exactly was said.”

It has three pillars, all from public content and traceable:

• 📚 Original Corpus (references/corpus/) — All of Zheng Xi’s public views from 2012–2026: investment operation analysis in periodic reports (quarterly / semi-annual / annual reports), fund manager notes, media interviews and reports, plus fund manager bio and lists of current/former funds managed.
• 🧭 Investment Method (references/method.md) — A methodological framework distilled from the above corpus, with each point backed by his original words. It allows the skill to extrapolate on topics not directly covered in the corpus, using Zheng Xi’s own methods, rather than saying “no record found.”
• 📊 Real Fund Data (references/fund_data/) — Real data snapshots of all 8 funds he managed (4 current + 4 former): quarterly top ten holdings, NAV/performance/size/asset allocation/tenure returns. Plus in references/all_funds/, a list of approximately 27,000 funds across the entire market, allowing on-demand fetching of any fund for comparison and scoring.

⚠️ For research and study assistance only, not investment advice.

🎯 What It Can Help You Do

💬 Examples · See How It Answers

Below are real outputs from the skill (slightly abridged).

1 Traceable Q&A — “What does Zheng Xi think about optical communications? When did he start favoring it?”

He explained the logic clearly in a June 2026 interview with China Securities Journal:

“Global AI capex has reached the trillion-dollar level. In the AI supply chain, data transmission is a crucial link, and optical communication is the key path for long-distance transmission. Combined with China’s global comparative advantage in optical communications, the market size for optical communication is expected to increase further.”

In terms of timing, his heavy position in “optics” can be traced back to Q2 2025. Cross-referencing his managed E Fund Information Industry Fund (001513) holdings: from 2025Q2 onward, optical communication names like Zhongji Innolight, Eoptolink, Yuanjie Technology, and Optowide Technologies collectively entered the top ten—words and actions match.

2 Direction Not Covered in Corpus — “Using Zheng Xi’s approach, what about innovative drugs?”

Zheng Xi has not publicly discussed innovative drugs in detail. The following is an extrapolation based on his consistent investment method and does not represent his personal views.

In all his public materials, “innovative drugs” appear only as a passing mention in quarterly review sections, never as a dedicated topic. So applying his method:

• Gate 1: Is there “supply-created demand”? Innovative drugs create previously non-existent demand through new targets/therapies—this fits his preferred “tech-driven inflation.” This gate aligns well.
• Gate 2: Where is China’s comparative advantage? Original innovation originates overseas; China’s advantage lies more in CXO / out-licensing BD—the “picks and shovels” part.
• Gate 3: Low ROE elasticity: Unprofitable biotech is a typical “pre-R&D low ROE” case, fitting his preference; but the liquidity gate would filter out thinly traded small-cap names.

(Involves specific companies/valuation—requires verification; not investment advice.)

3 Scoring Using Zheng Xi Framework — “Using Zheng Xi’s criteria, rate Zhaoshang CSI Baijiu. Would he buy it?”

Score for [Zhaoshang CSI Baijiu Index (012414)] Using Zheng Xi Framework: 21 / 100 · Not Aligned

One sentence: This is a qualified sector index tool, but almost nothing like “a fund Zheng Xi would buy.” Low score ≠ bad fund, just that the style is nearly opposite to his.

Dimension	Score	One-Sentence Basis
Direction / Inflation Attribute	4 / 25	Top ten are all premium baijiu, currently in a downward destocking cycle — “being priced down,” not “pricing up”
Low ROE Elasticity	6 / 20	All are long-term high-ROE baijiu blue chips, not his desired “low to be repaired”
Global Vision / China Comparative Advantage	2 / 15	Pure domestic consumption, unrelated to global tech cycles
Liquidity	7 / 10	Top holdings are all large-cap leaders with best liquidity (only bright spot)
Concentration & Cycle Splicing	2 / 15	Concentration 85.8% but passive lying flat, turnover proxy ~5% over recent 5 quarters—opposite to his “cycle splicing / high turnover”
Performance & Drawdown Verification	0 / 15	Latest 1Y −23.9%, max drawdown −63.2%, significantly underperforming

Would he buy it? Almost certainly not — a passively managed single traditional sector index fund completely mismatches his “top-down find inflation, prefer tech growth, buy low ROE elasticity, dynamic cycle splicing” method.

4 Full Market Comparison — “Compare Zheng Xi’s Information Industry with Ge Lan’s Zhongou Medical”

	Zheng Xi · Information Industry 001513	Ge Lan · Zhongou Medical 003095
Sector	Tech (Optical Communications · Compute · Semiconductors)	Healthcare (CXO · Innovative Drugs)
Style	Top-down, supply chain rotation, high turnover	Core assets long hold, low turnover
Latest 1Y	+290%	−5.4%
Max Drawdown	−46.3%	−63.9%

One sentence: Both are high-conviction single-sector managers, but Zheng Xi turns elasticity into returns by “riding the right cycle + dynamic rotation,” while Ge Lan’s “long hold core assets” fully absorbed deeper drawdowns during the healthcare downturn. (Data from public quarterly snapshots, as of 2026-06; not investment advice.)

⚖️ Differences from “Investment Framework/Methodology” Skills

🧩 Is It an Agent Skill?

Yes. It conforms to the Agent Skill specification: root directory contains SKILL.md (YAML frontmatter with name + description) + progressively loaded references/ and scripts/ resources, can be validated and packaged as a .skill file via official tools.

• Claude Code / Claude.ai natively support Agent Skill; place it in the skills directory and it loads automatically.
• Tencent WorkBuddy uses skill.yml manifest to organize Skills; this repo includes one, just put it in its skills/ directory (tested and working).
• In tools like ChatGPT / Gemini / Cursor that don’t have an “auto-load skill” feature, replicate using “custom instructions + knowledge files” — see Using in Other AI Tools.

🗂️ Directory Structure

zhengxi-views/
├── SKILL.md
├── skill.yml                      # Tencent WorkBuddy manifest (can ignore for Claude)
├── WORKBUDDY部署.md
├── README.md
├── references/
│   ├── method.md
│   ├── scorecard.md
│   ├── corpus_index.json
│   ├── corpus/
│   │   ├── 定期报告/
│   │   ├── 基金经理手记/
│   │   ├── 媒体报道/
│   │   ├── 简介.md
│   │   ├── 管理基金_在任.md
│   │   └── 管理基金_曾任.md
│   ├── fund_data/
│   │   ├── _index.md
│   │   ├── 001513_易方达信息产业混合/
│   │   └── ...(8 funds total, each containing 季度持仓.md and 净值业绩规模.md)
│   └── all_funds/
│       └── fund_list.json
└── scripts/
    ├── search_corpus.py
    ├── build_index.py
    ├── fetch_fund_data.py
    ├── build_fund_list.py
    ├── fund_lookup.py
    ├── fetch_any_fund.py
    └── score_fund.py

⚙️ Installation

Claude Code (Native, Recommended)

macOS / Linux:

mkdir -p "$HOME/.claude/skills/zhengxi-views"
cp -R SKILL.md README.md references scripts "$HOME/.claude/skills/zhengxi-views"/

Windows PowerShell:

$dst = "$HOME\.claude\skills\zhengxi-views"
New-Item -ItemType Directory -Force $dst | Out-Null
Copy-Item -Recurse SKILL.md,README.md,references,scripts $dst

After installation, fully restart Claude Code and ask “What does Zheng Xi think about optical communications?” to trigger the skill.

Dependencies (Only Needed for “Fetch Fund Data / Scoring”)

Pure corpus search does not require third-party libraries; to use fetch_* / score_fund for all-market data, install:

pip install -r requirements.txt  # requests / beautifulsoup4 / lxml

Reduce Confirmation Per Step (Claude Code, Optional)

The skill runs Python scripts. By default, Claude Code asks for confirmation each time. To skip confirmation, add a Python allow rule in ~/.claude/settings.json under permissions.allow:

{
  "permissions": {
    "allow": ["Bash(python:*)", "Bash(python3:*)"]
  }
}

(This allows all Python commands; recommended only on personal machines. Remove anytime. The script invocations are already designed as single commands without cd/redirection to minimize security prompts.)

🌐 Using in Other AI Tools

This project is an Agent Skill, natively supported by Claude ecosystem; other tools replicate via “instructions + knowledge files.” Features available depend on whether the tool can run Python and access the internet:

Tencent WorkBuddy runs locally and can run scripts with internet access, so all features (including real-time fetching) are available; it uses skill.yml as manifest, included in this repo. See Tencent WorkBuddy. ⚠️ = sandbox typically no external network, real-time fetch likely fails; you can run fetch_any_fund.py locally first to fetch the target fund, then upload the generated data file for offline scoring.

General approach (any AI that supports custom instructions + file uploads):

• Instructions: Paste the body of SKILL.md into “System Prompt / Custom Instructions.” The part about “how to run scripts” is Claude Code specific—can be removed or changed to “run uploaded scripts using code interpreter.” Keep behavior constraints (traceability, no fabrication, separate extrapolation from original words, bold disclaimer for out-of-corpus statements).
• Knowledge: Upload entire references/ (method.md, scorecard.md, corpus/, fund_data/, all_funds/fund_list.json); upload scripts/ if you need to run scripts.

Tencent WorkBuddy

WorkBuddy is a locally running AI workstation. Custom Skills are defined as a skill.yml manifest + implementation files (scripts) + README. This repo already includes skill.yml, usable directly:

1. Put the entire folder into WorkBuddy’s skills/ directory, enable it by importing via skill.yml (tested and working).
2. Run pip install -r requirements.txt to install script dependencies; because it runs locally, all features including real-time fetch and scoring are available.
3. Example trigger phrases: “Zheng Xi, what do you think about optical communications?” “Score Zhaoshang CSI Baijiu using Zheng Xi’s criteria.”

The skill.yml fields are written per general conventions; if WorkBuddy indicates mismatched fields, adjust slightly against its schema (the behavioral logic is all in SKILL.md). Full steps and alternatives in WORKBUDDY部署.md.

Cursor (IDE, Most Feature-Rich, Close to Claude Code)

1. Put the entire zhengxi-views/ folder into the workspace.
2. Create .cursor/rules/zhengxi.mdc and paste the body of SKILL.md as a Rule.
3. Cursor’s Agent has a local terminal, can run Python + internet—search, real-time fetch, scoring all work, usage same as Claude Code.

ChatGPT (Custom GPT)

1. Create a new GPT (My GPTs → Create).
2. Instructions: Paste the body of SKILL.md (adjust script-related wording as per “general approach” above).
3. Knowledge: Upload method.md, scorecard.md, corpus/ (can zip), fund_data/, all_funds/fund_list.json, and scripts/*.py.
4. Enable Code Interpreter (Data Analysis).
   • Available: Traceable Q&A, method, stylized commentary, Zheng Xi fund data, and script-based search/scoring on uploaded data.
   • Limited: Real-time full-market fetch (Code Interpreter has no external network)—instead, fetch locally first and upload.

Gemini (Gem / Google AI Studio)

1. Create a new Gem.
2. Paste SKILL.md body into the instruction bar; upload method.md, scorecard.md, corpus/, fund_data/ as knowledge.
3. Mainly supports reading features (Q&A / method / commentary / already packaged data cross-reference); no code + external network, so real-time fetch and script scoring should be done locally before asking.

Claude.ai (Web / Desktop)

• If your account supports uploading Skills, directly upload the packaged .skill file; or create a Project and upload SKILL.md with references/ and scripts/ as project knowledge.
• The web sandbox may lack internet access, so real-time fetch may be limited; built-in corpus and Zheng Xi fund data work normally.

🛠️ Script Quick Reference

search_corpus.py "optical communication"         Search corpus, return matching paragraphs + source (new→old)
search_corpus.py "ROE" "elasticity" --any        Match any keyword
fund_lookup.py 中欧医疗                           Look up fund code by name/code/type across market
fetch_any_fund.py 003095                         Fetch any fund's holdings/NAV/performance on demand to cache
score_fund.py 招商中证白酒                        One-click entry for Zheng Xi framework scoring (find code + prepare data + compute metrics)
build_index.py / build_fund_list.py               Rebuild after corpus / all-market list update

In Claude Code / Cursor, the AI will call these scripts automatically; you usually don’t need to type them manually.

✅ Data Sources and Authenticity

Corpus, investment method evidence, and fund data all come from publicly disclosed content by Zheng Xi and the funds he managed — periodic reports on E Fund’s official website, fund manager notes, media reports, and public fund data from Tian Tian Fund. All are real materials, traceable to original texts. The skill only searches and cites; it does not fabricate or invent.

🚧 Limitations

For research and study assistance only. Not investment advice. Does not predict prices, give buy/sell instructions, or guarantee returns. Citations must be faithful to the original text. The biggest taboo is making up statements he never said or rewriting original text to pass as his words — if unsure, go back to the corpus or honestly say “no record found.”

📄 License

MIT