After Serenity gained popularity, many people’s first reaction was to copy the setup. But we can actually equip Codex with these investment Skills and an AI investment research framework to turn it into a stock research assistant: one that can break down industry chains, read financial reports, gauge sentiment, check risks, perform valuations, write bearish reports, and even backtest strategies. Save these 5 GitHub projects first! Don’t let AI buy stocks for you directly; first let it help you do solid research.

• UZI-Skill 2.4K – The one Chinese investors should check out first! It integrates A-share, Hong Kong, and US stock analysis into a complete workflow: deep analysis, trap detection, Dragon and Tiger list, investment guru review panel, 22 dimensions, 180 quantitative rules. Suitable for individual stock screening, risk checks, and Chinese market research. GitHub: https://github.com/wbh604/UZI-Skill
• TradingAgents 83.7K – One of the most popular AI financial trading frameworks on GitHub. It splits an investment research team into multiple agents: fundamental analysis, sentiment analysis, news analysis, technical analysis, researcher, trader, risk management team. If you want to see how an “AI research team” collaborates, this project is the most representative. GitHub: https://github.com/TauricResearch/TradingAgents
• Serenity Skill 731 – An investment Skill widely discussed on X recently. It reuses Serenity’s core research path: starting from hot topics, breaking down the industry chain, finding supply chain bottlenecks, then returning to stock and fund directions, checking announcements, financial reports, customers, capacity, and risks. Suitable for researching directions with “long industry chains and messy information” like AI semiconductors, CPO, robots, computing power, power equipment, and innovative drugs. GitHub: https://github.com/muxuuu/serenity-skill
• Buffett Skills 512 – This Skill is for pouring cold water before you get excited. It forces you to check: Is this a good business? Does it have a moat? Is the cash flow reliable? Is management trustworthy? Is the price at a margin of safety? Suitable for long-term investors doing quality checks. GitHub: https://github.com/agi-now/buffett-skills
• QuantDinger 7.4K – A more complete AI quantitative trading workstation. It covers the full chain from AI research, strategy code, backtesting, simulation trading, to live execution and monitoring. If you’re not satisfied with “letting AI analyze stocks” and want to turn ideas into strategies and validate them, QuantDinger is worth further study. GitHub: https://github.com/brokermr810/QuantDinger

Don’t let AI buy stocks for you. Let Codex break down industry chains, read financial reports, check evidence, find risks, write bearish reports, and do backtesting. That’s the usage ordinary investors should copy most. #codex #skills #Serenity

wbh604/UZI-Skill

Source: https://github.com/wbh604/UZI-Skill

UZI-Skill

“51 investment gurus help you watch the market; Buffett and Zhao Laoge finally sit at the same table.”

License: MIT

Python 3.9+ (https://python.org) Claude Code (https://claude.com/product/claude-code)

Dimensions Investors Methods Self-Review

A-shares / Hong Kong / US stocks · Individual stock deep analysis engine · v3.7.1 Homepage update: Serenity (AI positioning/bottleneck hunter) introduction + --school H/I released · v3.7.0 13 new tech big shots joined (a16z / Naval / Jensen Huang / Musk / Gao Ling Zhang Lei / Burry / Chanos · 52→65 judges) · v3.6.2 cninfo pagination long tail fix

Installation · Usage · Three Depth Levels · Hermes 🆕 · Review Panel · Serenity 🆕 · Institutional Methods · Self-Check Gate · Report Screenshots · FAQ · Join Communication Group for Testing · Contributors

中文 | English

🚀 30-Second Quick Start

Drop a single command in any agent · Works after installation. Detailed installation see Installation.

Most commonly used 4 commands after installation (say directly in any agent):

/stock-deep-analyzer:analyze-stock 贵州茅台 ← Full 22-dimension × 65-judge analysis (5-8 min)
/stock-deep-analyzer:quick-scan 002217 ← 30-second quick check
/stock-deep-analyzer:scan-trap 002217 ← Pump-and-dump check
/stock-deep-analyzer:dcf 600519 ← DCF valuation special

💡 Current stable version v3.3.2:

• v3.3.2 · 🆕 GitHub issue #50 + #51 hotfix (community-driven) · #50 institutional missing import svg_sparkline causing Stage 2 NameError · #51 XueQiu cubes_search.json endpoint deprecated → switched to query/v1/search/cube/stock.json (thanks @chenxiang-bj / @bilieebiliee1-design / @Kylin824)
• v3.3.1 · Hermes compatibility regression fix (v3.x refactoring period main branch missing hermes adaptation · community reported root cause) · main branch now directly supports hermes skills install
• v3.3.0 · Branch consolidation · segmental rendering layer cherry-picked to v3.2 architecture · 22 stale branches cleaned (only main + hermes-compat retained)
• v3.2.0 · assemble_report.py 2964 → 587 lines (-80%) · Split into 5 lib/report/*.py submodules
• v3.1.0 · run_real_test.py 2105 → 735 lines (-65%) · 1228 lines of pure functions moved to lib/pipeline/score_fns.py
• v3.0.0 · Pipeline architecture enabled by default (python run.py now uses new path · UZI_LEGACY=1 falls back to old path)

Two huge files total 5069 → 1322 lines (-74%) · 332 tests all pass · Real-device e2e 002217 resume generates report in 10s · All v2.x APIs 100% backward compatible.

v2.15 series retained: capital_flow universe cache (100x speedup) · school_scores by school scoring · Hybrid formula + polarization stretch.

Hermes users with old version remnants may encounter errors · Reinstalling once will fix (hermes skills uninstall then install 4 skills) · See INSTALL-HERMES.md.

💬 Testing and Communication Group

Group invites get blocked. If you have ideas, add me directly. Only for the plugin itself or other interesting projects and quantitative topics, not individual stocks.

Too many people adding, please include the matter in your note.

Acknowledgements

Learn AI, go to L! Thanks to the Linux.do (https://linux.do/) community support.

What is this?

In one sentence: input a stock, Claude becomes your personal analyst, runs through data across 22 dimensions, calls 17 Wall Street analysis models, lets 51 entirely different investment gurus score, and finally outputs a 600KB Bloomberg-style report.

/stock-deep-analyzer:analyze-stock 国盾量子

After 5-8 minutes you’ll get:

• An HTML report — can be opened directly in a browser, self-contained, viewable offline
• A WeChat Moments vertical image — 1080×1920, ready to share
• A WeChat group battle report — 1920×1080
• A one-paragraph summary — copy and paste into group chat

Why build this?

Previously, to research a stock: check fundamentals on East Money → look at K-lines on Tonghuashun → browse what big V’s say on Xueqiu → find sell-side reports in research systems → calculate DCF in Excel → still buy and lose money.

These tasks are essentially “gather information → multi-angle analysis → draw a conclusion”. Why not let AI do it all?

After looking around, most offerings are either GPT wrappers outputting three paragraphs of fluff or unaffordable institutional terminals. Anthropic released financial-services-plugins (https://github.com/anthropics/financial-services-plugins) with good methodology (DCF / Comps / LBO), but it’s entirely US stock oriented + requires paid data sources.

So I built my own. All free data sources, zero API key, works directly for A-shares.

Installation

No matter which agent you use, just drop one command:

Claude Code

/plugin marketplace add wbh604/UZI-Skill
/plugin install stock-deep-analyzer@uzi-skill

After installation, say /stock-deep-analyzer:analyze-stock 贵州茅台.

⚠️ Must include the stock-deep-analyzer: namespace prefix

After installing the plugin in Claude Code, all skills/commands are prefixed with stock-deep-analyzer:. In some environments, the short name (/analyze-stock) won’t be auto-resolved — for safety always use the full name:

• /stock-deep-analyzer:analyze-stock
• /stock-deep-analyzer:quick-scan
• /stock-deep-analyzer:scan-trap
• /stock-deep-analyzer:dcf
• /stock-deep-analyzer:ic-memo
• /stock-deep-analyzer:investor-panel
• /stock-deep-analyzer:trap-detector
• /stock-deep-analyzer:deep-analysis
• … and all 14 commands

Similarly for Cursor / Gemini CLI / Codex: always use /stock-deep-analyzer: full name to avoid short name resolution failure.

Codex

Simply say to Codex:

Please install UZI-Skill following the instructions at https://raw.githubusercontent.com/wbh604/UZI-Skill/main/.codex/INSTALL.md, then do a deep analysis of 贵州茅台.

OpenClaw / Lobster

Say to Lobster:

Install this stock analysis skill from https://github.com/wbh604/UZI-Skill, then analyze 贵州茅台.

Cursor

/add-plugin stock-deep-analyzer

Then say “analyze 贵州茅台”.

Gemini CLI

gemini extensions install https://github.com/wbh604/UZI-Skill

OpenCode

Say to OpenCode:

Please install and analyze 贵州茅台 following https://raw.githubusercontent.com/wbh604/UZI-Skill/main/.opencode/INSTALL.md.

Windsurf / Devin / Other Agents

Drop this command:

Clone https://github.com/wbh604/UZI-Skill, read AGENTS.md to learn how to use it, and perform a deep analysis of 贵州茅台.

📱 Not at your computer?

Say to any agent:

Analyze 贵州茅台 in remote mode, generate a public link so I can view it on my phone.

The agent will automatically start a Cloudflare Tunnel with --remote and give you a https://xxx.trycloudflare.com link.

Usage

Full Deep Analysis (5-8 minutes)

/stock-deep-analyzer:analyze-stock 水晶光电
/stock-deep-analyzer:analyze-stock 002273
/stock-deep-analyzer:analyze-stock 00700.HK
/stock-deep-analyzer:analyze-stock AAPL

Specialized Commands

All require the /stock-deep-analyzer: prefix to ensure execution.

🎯 Scoring Calibration (v2.11)

User feedback: “Moutai scored 47”, “Never exceeds 65” — diagnostics found two formula strictness issues. v2.11 calibration:

Formula (unchanged): overall = fund_score × 0.6 + consensus × 0.4 Typical white horse (e.g., Moutai) expected:

• v2.9.1: fund=62 consensus=45 → overall 55 → priority watch
• v2.11: fund=62 consensus=50 → overall 57 → priority watch (but closer to “worth watching” boundary, easier to reach 65 when white horse trend starts)

Total impact across both tiers ~5-8 points. Real traps still < 35 → avoid, score discrimination actually improved.

Diagnostic field panel.json::consensus_formula.version = "v2.11 · (bullish + 0.6*neutral) / active" is auditable.

Regression tests: tests/test_v2_11_scoring_calibration.py 8 test cases.

Full calibration record see BUGS-LOG.md v2.11.0 section.

🎚️ Three Thinking Depths (v2.10.3 New)

Let users choose analysis intensity — quick / normal / deep:

python run.py 600519 --depth lite   # ⚡ Quick mode (1-2 min)
python run.py 600519                # 📊 Standard analysis (5-8 min) · Default
python run.py 600519 --depth deep   # 🔬 Deep research (15-20 min)

Or via environment variable:

export UZI_DEPTH=lite   # or medium / deep
python run.py 600519

Comparison Table

Auto-Degradation Strategy

• First installation / .cache/_global empty → automatically switch to lite (reduce first cold start time)
• Network pre-check 3+ domains unreachable → automatically switch to lite (avoid getting stuck)
• Manual --depth always overrides auto-detection

Practical Choices

Command Mapping (Implicit Levels)

🎭 65 Judges

Not template talking points. Each has their own quantified rule set (236 total), and their advice must cite which specific rules were triggered.

As of v3.7.0, 13 new tech big shots + an independent Group I Serenity (AI positioning/bottleneck hunter) added, covering 9 major schools:

v3.7.0 (2026-06) new 13 people bolded. Group H are tech CEOs with their own industry chain perspective; Group I Serenity is a single special judge (see below).

Example:

Buffett scores 水晶光电 62 · Neutral “Watch: moat 27/40 visible; but ROE 5-year low 6.7%, only 0/5 threshold met” ✅ Asset-liability ratio 30% conservative · ❌ ROE 5-year low 6.7%

Jensen Huang scores a certain CPO optical module stock 100 · Bullish “On the AI compute chain · Direct beneficiary of data center CapEx · Gross margin ≥50% pricing power strong — this is a beneficiary of Moore’s Law for optics.” ✅ Strong correlation with AI compute demand · ✅ Deep CUDA/ecosystem binding

Klarman scores 水晶光电 0 · Bearish “Core bearish: No 30% margin of safety”

🧠 Group I · Serenity · AI Positioning/Bottleneck Hunter

Major character: @aleabitoreddit (https://x.com/aleabitoreddit), a retail researcher who went viral on X (Twitter) in 2026. Standalone group with standalone scoring — because her approach is extremely concentrated and contrarian, unlike any institutional big shot.

Who is she?

• Self-described background: Former AI research scientist · Nature paper author · Former RISC-V Foundation member · Semiconductor/optical communication engineer
• Anime avatar, anonymous, no face reveal, doesn’t sell courses, doesn’t copy trade, research all free and public, 300K+ X followers
• Claimed victory: Predicted the InP substrate bottleneck stock AXTI (12 → $70+, peak $115–140) about a year early, confirmed by IntelliEPI CEO in 2026 Q1 as “InP shortage is the bottleneck of entire AI infrastructure”

⚠️ Identity and returns are self-reported/media-reported, not audited by third parties, with conflicting numbers across sources. This project only distills her methodology as one analytical perspective, not an endorsement of her actual track record. See docs/serenity-research-dossier.md (archive of 20+ sources online).

Her role in UZI-Skill

Her “AI supply chain chokepoint/bottleneck theory (Chokepoint Theory)” is built into a quantifiable judge — Don’t buy AI leaders (stocks like NVIDIA that are already fully priced), but go upstream along the supply chain to find that “global dependency that’s most likely to hit supply limits” — overlooked small/mid-cap upstream players, and position before the market prices it in.

Leader bought to the moon → Go upstream → Find the hardest-to-replace link → Find the tightest-supply small cap in that link → Position early

Core scoring logic: “Position determines stance” — ignore cheap valuation, ignore fast growth, focus on one variable: Does this company’s product choke others in the current AI wave?

To judge if a link is a “chokepoint” look at three things: 1 Hard to replace (how long to switch supplier/material/process, longer is better → 14_moat switching cost) 2 Tight supply (capacity can’t keep up with AI demand curve, worse is better → 7_industry) 3 Not yet priced (market still sees it with old narratives like “cyclical stock / old semiconductor / niche material” → 5_chain + 15_events).

How to run Serenity perspective alone

python run.py 300394.SZ --school I   # Only Serenity's "chokepoint or not" judgment
python run.py NVDA --school H        # Only Group H tech leader school (Jensen Huang/Musk/Altman/Saylor)

Methodology six-step + alpha 5 dimensions see skills/deep-analysis/references/fin-methods/serenity-bottleneck.md; Tone library + scoring rules see skills/investor-panel/references/group-i-serenity.md.

📐 17 Institutional Methods

Methodology ported from anthropics/financial-services-plugins (https://github.com/anthropics/financial-services-plugins), adapted to A-share parameters (rf=2.5% / ERP=6% / tax rate 25% / terminal g=2.5%):

Valuation Modeling

• DCF (WACC decomposition + two-stage FCF + Gordon Growth terminal value + 5×5 sensitivity heatmap)
• Comps peer comparison (PE / PB / EV-EBITDA percentiles + implied target price)
• Three-statement model (5-year IS / BS / CF linkage)
• Quick LBO (PE fund perspective IRR cross-check)
• Merger accretion/dilution model

Research Workflows

• Initiation report (JPM/GS/MS format · rating + target price + thesis + risks)
• Earnings beat/miss interpretation
• Catalyst calendar (real event extraction + future schedule + impact rating)
• Investment logic tracking (5-pillar health)
• Morning note · quantitative screen · industry overview

Deep Decision Making

• IC investment committee memo (8 sections · Bull/Base/Bear three scenarios)
• Porter’s Five Forces + BCG matrix
• DD due diligence checklist (5 workflows, 21 items · auto-mark completion status)
• Unit economics · value creation plan · portfolio rebalancing

📸 What Reports Look Like

All screenshots below are from real analysis of 水晶光电 (002273.SZ).

Overall Score + Key Conclusions

The Great Divide: Bull vs Bear

Fisher 100 vs Klarman 96, three rounds of arguments, each citing specific numbers.

65 Judges · Tribunal Panel

Each person a light — green bullish, red bearish, gray neutral.

Chat Room Mode

Judges speak in their own language style, citing the specific rules triggered.

DCF Valuation · 5×5 Sensitivity Heatmap

WACC 6.96% · Intrinsic value ¥20.73 · Margin of safety -28.6%, color from dark green (undervalued) to dark red (overvalued).

IC Investment Committee Memo · Three-Scenario Returns

Bull ¥26.95 / Base ¥20.73 / Bear ¥14.51, each with probability and assumptions.

22-Dimensional Deep Card

Each dimension has independent visualization — K-line candlestick / PE Band / radar chart / supply chain flowchart / thermometer / donut chart.

WeChat Moments Vertical Image · One-click Share

🔧 Data Sources

All free, zero API key:

Multi-layer fallback chain — if one source fails, automatically switch to next.

📱 6-Platform Social Hot Lists (v2.12 new)

Retail sentiment and pump-and-dump themes often surface first on Douyin/Xiaohongshu/Weibo, which DuckDuckGo can’t scan. From v2.12, the 17_sentiment dimension automatically checks:

• Weibo Hot Search · fetches weibo.com/ajax/side/hotSearch · 50 real-time hot topics
• Zhihu Hot List · fetches zhihu.com/api/v3/feed/topstory/hot-list-web · 50 items
• Baidu Hot Search · fetches top.baidu.com/api/board · real-time list
• Douyin Hot Topics · fetches douyin.com/aweme/v1/web/hot/search/list/ · search hot topics
• Toutiao Hot List · fetches toutiao.com/hot-event/hot-board/ · hot events
• Bilibili Hot Search · fetches s.search.bilibili.com/main/hotword · site-wide hot words

Stock names (including short names, e.g., “贵州茅台” → “贵州”/“茅台”) matched in hot list titles → counted into sentiment heat + specific entries recorded.

Data structure: synthesis’s 17_sentiment.data.hot_trend_mentions:

{
  "stock_name": "贵州茅台",
  "platforms_ok": 6,
  "total_hits": 3,
  "by_platform_count": {"weibo": 2, "zhihu": 1, ...},
  "mentions": {
    "weibo": [{"rank":3, "title":"茅台 1499 回归", ...}],
    ...
  }
}

Acknowledgement: This module’s design referenced the hottrend service implementation from run-bigpig/jcp (https://github.com/run-bigpig/jcp) (Jiucai Pan AI).

🔑 Optional: East Money Miaoxiang Skills API (v2.3 new)

In 2026, push2.eastmoney.com is frequently blocked by anti-scraping on mainland networks. If MX_APIKEY is set, UZI-Skill will prefer the official NLP API:

• Chinese name correction: “北部港湾” → auto-recognized as “北部湾港(000582.SZ)”
• Quote snapshot: bypass push2 to get latest price/market cap/PE/PB/industry

Configuration:

cp .env.example .env
# Edit .env and fill MX_