UZI-Skill: A 66-Investor Jury for A-share, HK and US Stock Analysis Inside Your Coding Agent

The workflow UZI-Skill is trying to replace

The README opens with a confession that reads like a support-group introduction: the old routine was scrolling East Money for fundamentals, switching to Tonghuashun for candlesticks, checking Xueqiu for what large accounts said, hunting sell-side notes, and hand-building a DCF in Excel, only to lose money anyway. UZI-Skill compresses that loop into one command against one ticker. The target user is a retail or semi-professional investor who already runs an agent-based coding environment and is comfortable with a terminal. The project is explicit that it covers individual stocks only; active funds and fund-manager behaviour belong to a sibling repository, jilao-skills, which the README links at the top. That split matters, because it tells you the scope is deliberately narrow: one ticker in, one report out.

What actually happens between the ticker and the HTML file

The pipeline is described as the default backbone from v3.0 onward, with the older v2.x code path still reachable through the UZI_LEGACY=1 environment variable. Data collection pulls from free, keyless sources: akshare for A-share and HK, yfinance for US, and baostock as a candlestick fallback. The requirements file pins baostock at 0.9.1 or newer with a comment explaining that the server side began requiring that version on 2026-04-22 and older clients fail at login(). exchange-calendars is constrained to versions below 5.0 for trading-day and holiday validation across the three markets.

On top of that data layer sit the scoring layers. The README describes 22 data dimensions, 22 institutional methods (DCF, comps, LBO, initiation notes, IC memos and more), and a jury of 66 investor personas grouped into nine schools, with an independent I group called Serenity aimed at AI supply-chain bottlenecks. A self-review module at skills/deep-analysis/scripts/lib/self_review.py runs 13 mechanical checks before a report is emitted, which is the project's own answer to the obvious failure mode of an LLM writing confident nonsense. The self-review gate is the most interesting design decision here: instead of trusting the model's prose, the pipeline verifies structural conditions and blocks output that fails them.

Installing UZI-Skill in Claude Code and running a first analysis

The README gives a marketplace route for Claude Code. Two commands register the plugin and install it, and the README warns that the stock-deep-analyzer: namespace prefix is required because short names like /analyze-stock do not resolve reliably in every environment.

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

After installation, the full analysis command takes a Chinese name, a numeric A-share code, an HK code with the .HK suffix, or a US ticker. The README states the run takes five to eight minutes and produces a self-contained HTML report, a 1080x1920 vertical image, a 1920x1080 group-chat card, and a paste-ready text summary.

/stock-deep-analyzer:analyze-stock 贵州茅台
/stock-deep-analyzer:analyze-stock 002273
/stock-deep-analyzer:analyze-stock 00700.HK
/stock-deep-analyzer:analyze-stock AAPL

If you would rather skip the plugin system, the README's CLI path clones the repository, installs Python dependencies, and calls run.py directly with a stock name.

git clone https://github.com/wbh604/UZI-Skill.git && cd UZI-Skill && pip install -r requirements.txt && python run.py 贵州茅台

A faster first look is quick-scan, which the README describes as a 30-second judgement, and scan-trap, which screens for pump-and-dump patterns. Both accept the same ticker formats.

The optional East Money key and what breaks without it

The .env.example file documents one recommended credential, MX_APIKEY, obtained free from a link in the file. Setting it changes two things: Chinese-name correction, so a misspelling like 北部港湾 is resolved to 北部湾港, and a quote snapshot that the file says is far more stable than scraping push2.eastmoney.com. That phrasing is the project admitting its default path is a scraper. Scrapers break. If you run without the key, name typos will not be corrected and quote retrieval depends on an endpoint the maintainer describes as less stable. Other environment switches are STOCK_NO_CACHE=1 to bypass all API caching, UZI_NO_AUTO_OPEN=1 to stop the browser from opening after analysis, UZI_DISABLE_GLOBAL_PEERS=1 to keep only local-market peers, and UZI_GLOBAL_PEER_LIMIT with a documented range of 3 to 12 for global peer completion.

Where the 66-investor jury is weaker than it sounds

The headline number is the weakest part of the pitch. A jury of 66 named investors, from Buffett to 赵老哥 to a persona called 股海贼王, is a scoring rubric, not 66 opinions. The README itself says the v3.9.0 addition of 股海贼王 was the first judge distilled from real settlement statements, which implies the earlier 65 were built from something else, presumably public style descriptions. A judge distilled from a description and a judge distilled from trade records are not the same kind of object, and the report does not appear to distinguish them at scoring time. The Serenity group is more defensible because it ships with an explicit evidence ladder: the README describes three tiers where a claim of verified mass production scores around 90 and a pure theme claim scores around 60, plus eight penalty factors and an eight-layer supply-chain breakdown. That is a mechanism you can inspect. The rest of the jury is a mechanism you largely have to take on faith.

When to reach for something else

The README points at Anthropic's financial-services-plugins as the closest methodological relative and names the difference directly: that project is US-market oriented and assumes paid data sources, while UZI-Skill is built for free sources and runs on A-shares out of the box. If your universe is US large caps and you already pay for a data feed, the Anthropic plugin's DCF and comps tooling is the more conventional choice. If your question is about fund managers rather than stocks, the README redirects you to jilao-skills, which evaluates buying and selling behaviour from public disclosures rather than net asset value. And if you need a number you can defend to a compliance officer, none of these are the right tool: UZI-Skill is a research accelerator, and its own self-review gate exists because its output is generated text.

Maintenance, releases and the MIT licence

The repository is not archived and the last push was on 2026-09-05. Releases are frequent and granular: v3.9.1 on 2026-06-22 made the report navigation bar collapsible, v3.9.0 on 2026-06-11 added the 股海贼王 judge, and v3.8.1 on 2026-06-10 was a consistency pass filling six gaps across the H and I layer pairs. package.json reports version 3.9.4 while the README's badge line also says v3.9.4, so the two agree even though the release list stops at 3.9.1. The Python floor is 3.10, stated in the README badge. Upgrade cost is mostly dependency drift, and requirements.txt shows the project already absorbing that: the baostock comment records a server-side break in April 2026 that forced a minimum version, and exchange-calendars carries an upper bound below 5.0. The MIT licence is permissive and places no conditions on commercial use beyond attribution; the repository carries no separate terms for the bundled investor personas, and nothing in the README addresses whether naming real investors in a scoring rubric raises issues in any jurisdiction. That is a question for your own counsel, not for the README.

Editorial conclusion

Adopt UZI-Skill if you already work inside Claude Code, Codex, Cursor or Gemini CLI and want a structured, offline-readable second opinion on a single ticker, and you accept that the 66 personas are scoring rules rather than real money. Do not adopt it if you need audited financials, intraday signals, or a portfolio system that survives a data-source outage, because the pipeline leans on free scraped endpoints and a single optional API key. Verify first that your Python is 3.10 or newer, that pip install -r requirements.txt resolves akshare and baostock, and that playwright install chromium actually downloads a browser, since the share-card and war-report screenshots depend on it.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. wbh604/UZI-Skill on GitHub