# Fund Guy Skill: Behavioural Audit Engine for Chinese Mutual Fund Managers > Fund Guy Skill is a Python and AI agent workflow that audits a Chinese mutual fund manager's actual trade decisions against their outcomes, producing a single-file interactive HTML report with scoring, K-line battle replay, buy-sell autopsies, independence testing, and star-manager pattern detection. It requires only public data and no API keys. **wbh604/fund-guy-skill** — 糟糕，我被基佬包围了！那么这个时候就有人要问了，主播主播，有没有什么简单好用的基佬筛选办法？有的兄弟，有的，快来看看jilaoskill吧！ - Repository: https://github.com/wbh604/fund-guy-skill - Website: https://wbh604.github.io/jilao-skills/ - Stars: 571 · Forks: 59 - Language: HTML - License: MIT - Published: 2026-09-17 · Updated: 2026-09-17 · Language: en - Canonical page: https://hysenlabs.com/projects/wbh604-fund-guy-skill ## Why Trade Behaviour Matters More Than NAV Returns Most fund evaluation tools in China show net asset value curves and platform ratings. Fund Guy Skill takes a different position: a manager's NAV gain may reflect style tailwinds rather than skill, and platform scores typically do not verify the timing of individual sell decisions. The README summarises this: "The platform gives him 84 points, we pull every trade and give him 62." The tool addresses four questions the README identifies as typically unasked: 1. NAV gains may come from market beta rather than manager decisions. 2. Platform ratings measure returns and tenure, not whether sell decisions protected capital. 3. A manager who differs from peers may still be losing money when he diverges. 4. Copying a manager by reading quarterly reports may recover most of the alpha after the first-mover advantage is priced in. The audit engine runs on Chinese fund code 163417 (Xingquan Heyi) as the demo, with the manager's name and photo anonymised in the public example. The methodology is documented in 18 rules in `skills/fund-manager-alpha/SKILL.md`. ## Data Sources and Zero API Key Requirement All data comes from free public sources with no API keys needed. The README lists the data and its origin: NAV, scale, subscriptions/redemptions, holder data, manager records, and platform scores come from Tiantian Fund (East Money). Quarterly holdings and fund rankings come from East Money via akshare. Stock industry classifications come from East Money F10 and Hong Kong F10. Management and custody fees come from East Money's fund fee pages. Full-market holdings cross-sections come from cninfo via akshare. A-share weekly K-lines (adjusted for splits) come from baostock. Hong Kong weekly K-lines come from Sina Finance via akshare. Subscription gates (purchase limits, openings, company self-purchases) come from East Money fund announcements. Other products under the same manager come from East Money via akshare. All raw data is written to `.cache/` with interface name, parameters, and fetch timestamps. Any number that cannot be traced to a source is not allowed in the report. ## Installing and Running the Audit The README documents several ways to run Fund Guy Skill depending on the AI client in use. For a direct command-line run: ```bash python3 -m venv .venv && .venv/bin/pip install -r requirements.txt ``` Then fetch the data for fund 163417 and run the analysis: ```bash .venv/bin/python scripts/fetch_fund.py 163417 .venv/bin/python scripts/fetch_stock_klines.py 163417 .venv/bin/python scripts/fetch_stock_industry.py 163417 .venv/bin/python scripts/fetch_pingzhong.py 163417 .venv/bin/python scripts/fetch_house.py 163417 .venv/bin/python scripts/fetch_market_similar.py 163417 ``` The dependencies are minimal: akshare 1.18 or later, baostock 0.8 or later, pandas, and requests, as listed in `requirements.txt`. For Claude Code users, the README documents a plugin install path: ```bash /plugin marketplace add wbh604/fund-guy-skill /plugin install fund-manager-alpha@fund-guy-skill ``` This exposes three slash commands: `/analyze-fund 163417` for a full audit, `/quick-check 163417` for a five-minute check, and `/find-similar 163417` for holdings overlap analysis. ## What the Report Contains The output is a self-contained HTML file of approximately 700 KB that works offline. The report includes: A verdict with a total score (example: 62 out of 100) broken into timing weight (35%), control weight (35%), and alpha quality weight (30%), each with rules the reader can verify. A comparison with the platform's score and the documented reason for the gap. A K-line battle replay showing the manager's actual weekly K-lines with buy and sell points marked. Green dots indicate buys, red dots indicate active sells, and yellow dots indicate passive reductions from hitting the 10% concentration limit or facing redemptions. The animation replays the year week by week. A buy-sell autopsy for every decision: in the demo, 65 buy decisions and 16 full exits were each verified against 12-month forward performance. The buy win rate and the capital-protection rate from exits are shown separately. An independence audit comparing the manager's holdings to peers at the same fund house, and verifying whether periods of high divergence outperformed the market. A star-manager pattern check covering nine suspicious patterns such as cherry-picking at fund launch, hiding poor performance in merged products, and opening subscriptions at market tops. Patterns that could not be verified are marked "unverified" rather than "absent". ## Three Report Style Templates Beyond the main report, the repository ships three alternative visual styles for the same methodology and data. The README describes them: Style A presents the report as an investigative dossier with a kraft-paper-and-official-stamp aesthetic, treating the manager as a subject under investigation with personal files and an evidence section. The demo is at `assets/style-a-dossier.html`. Style B uses trading card game language, presenting rarity, attribute panels, combat records, and fatality replays. The demo is at `assets/style-b-card.html`. Style C uses medical checkup language, with a summary conclusion, test results, prior history, and dosage recommendations. The demo is at `assets/style-c-checkup.html`. All three are rendered from mock data as style templates. The main report method uses real fetched data. ## Methodological Rules and What Cannot Be Automated The 18 methodological rules in SKILL.md govern what the tool can and cannot claim. Several directly constrain the output: Beta cannot be counted as alpha. Style-driven returns (market cap factor, growth-value factor) must be stripped before scoring the manager's contribution. Only active full exits count as sells for capital protection analysis. Passive reductions from hitting the 10% concentration rule or from fund redemptions are excluded from sell-decision scoring. Failed independent judgements must all be recorded. Reporting only winning divergences invalidates the independence audit. Numbers that cannot be computed are labelled "unverified" rather than inferred. The announcement full text and mandate analysis are left to the AI agent rather than automated scripts. The nine star-manager checks that can be filtered from structured data use real fetched data; those that cannot are marked as unverified. The README notes that the scripts are a reference implementation for one real run. If a different fund is substituted, the K-line window, industry classification, fee lookup, style index, peer group, and gate analysis must be recomputed for that fund. ## Limitations and Maintenance Fund Guy Skill covers only mainland Chinese and Hong Kong-listed funds sourced through the East Money ecosystem. It does not support funds listed on other exchanges or managed outside China. The dense routing and automated brokerage analysis does not cover equity funds with fewer than 8 years of decision history, as the methodology requires sufficient trade data to compute statistically meaningful ratios. The report opens correctly only in a browser with JavaScript enabled. The README explicitly warns against previewing the HTML in htmlpreview.github.io, where embedded scripts often fail to run, making scores appear as zero and K-lines appear empty. The last push to the repository was on 2026-09-03. The repository is not archived. The licence is MIT, permitting use in commercial applications. The TradingView Lightweight Charts library used in the K-line display carries its own Apache-2.0 licence documented in the README's data sources table. ## Conclusion Fund Guy Skill is the right tool for Chinese retail investors and quantitative hobbyists who want a verifiable, data-backed review of a fund manager's decision quality before investing. It is not designed for institutional analysis that requires real-time data feeds or programmatic integration into a portfolio management system. The output is a standalone HTML file; there is no database, API, or scheduled refresh. Before running it on a fund other than the demo 163417, read the SKILL.md methodology to confirm the three-tier data model and the 18 methodological rules fit the fund you are auditing. ## FAQ ### Does Fund Guy Skill require paid data subscriptions or API keys? No. The README states all data sources are free and no API keys are required. Data comes from East Money, baostock, Sina Finance via akshare, and cninfo via akshare. ### Can Fund Guy Skill analyse any Chinese mutual fund, not just the demo fund 163417? Yes, any fund with a code can be analysed by substituting the fund code in the fetch scripts. The README notes that when switching funds, the K-line window, industry data, fee lookups, style indices, peer comparison group, and gate analysis must all be recomputed for the new fund; the hardcoded demo parameters for 163417 must not be carried over. ### What AI clients does Fund Guy Skill support for agent-driven analysis? The README documents installation paths for OpenClaw (Lobster), Claude Code via plugin or skill, Codex, Cursor, Windsurf, Devin, and other agents through the AGENTS.md general guide. A plain command-line path using Python scripts is also available for clients not listed. ## Sources - [Issues](https://github.com/wbh604/fund-guy-skill/issues) - [License: MIT](https://github.com/wbh604/fund-guy-skill/blob/main/LICENSE) - [Project website](https://wbh604.github.io/jilao-skills/) - [README](https://github.com/wbh604/fund-guy-skill/blob/main/README.md) - [wbh604/fund-guy-skill on GitHub](https://github.com/wbh604/fund-guy-skill) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/wbh604-fund-guy-skill