fund-guy-skill: a behavioral audit engine for Chinese fund managers

The problem: platform scores never check the sell side

Most retail fund evaluation reduces to two inputs: the NAV curve and the platform score. The README states the project's complaint plainly: a rising NAV does not prove the manager is skilled, because a broad growth style in 2020 would have carried almost anyone. Platform scores, it argues, are built from return and years of tenure, and nobody verifies whether the manager sold at the right time. The README also takes aim at the independence claim, noting that being different from the house book is praised without checking whether the divergent bets made money.

The intended user is someone who already invests in Chinese public funds and wants a decision record rather than a marketing number. The README's own framing is blunt: the platform gives a manager 84, the skill gives him 62, and the 22-point gap comes from the platform not looking at when he sold.

How the audit works: fetch, analyze, render one HTML file

The pipeline is a sequence of standalone Python scripts driven by a fund code, with the agent runtime acting as the orchestrator. The README splits the work into three stages. Fetch scripts pull primary data (NAV, holdings, manager, holders, fees), weekly K-lines for the heavy holdings, industry classification, the Tiantian Fund pingzhongdata payload (subscription and redemption data, platform score, manager photo), peer-fund holdings, a market-wide overlap scan, the year's top-ten funds, and subscription gate announcements. Analysis scripts then run the behavioral work: trade autopsy and scoring, divergence from peers, and the gate-by-market-percentile check. A final build script emits a single self-contained HTML file, roughly 700KB according to the README, that opens offline.

The scoring model is stated as a total of 62 for the 163417 example, composed of timing at 35%, control at 35%, and excess quality at 30%. The autopsy verifies each action against the price 12 months later: the README reports a 58% hit rate on buys (beating the index counts as a win) and a 31% rate on exits avoiding declines, summarized as "buys well, sells badly." A factor decomposition splits 103 months of returns across market, size, and growth-value, leaving +7.7% per year as the residual. The README is explicit that unclear items are marked unverified rather than written up as absent.

Installing fund-guy-skill and running a first analysis

The simplest path is the OpenClaw route, which the README describes as a single command that clones the repository into the skills directory and installs dependencies:

git clone https://github.com/wbh604/fund-guy-skill.git ~/.openclaw/skills/fund-guy-skill && pip install -r ~/.openclaw/skills/fund-guy-skill/requirements.txt

After that you address the agent in natural language, pointing it at skills/fund-manager-alpha/SKILL.md and the reference scripts under scripts/. The dependencies are listed in requirements.txt as akshare>=1.18, baostock>=0.8, pandas, and requests, so the install is four packages and no API key.

For a plain command-line run, the README's quick-start creates a virtual environment, installs requirements, and then chains the fetch and analysis scripts for fund 163417:

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python scripts/fetch_fund.py 163417
.venv/bin/python scripts/fetch_stock_klines.py 163417
.venv/bin/python scripts/analyze_fund.py 163417
.venv/bin/python scripts/build_fund_report.py 163417

The full quick-start in the README runs eleven scripts in order: fetch_fund, fetch_stock_klines, fetch_stock_industry, fetch_pingzhong, fetch_house, fetch_market_similar, fetch_top_funds, fetch_gates, then analyze_fund, analyze_house, analyze_gates, analyze_addons, and finally build_fund_report. The README's condensed Codex instruction is to run `python run.py 163417 --no-browser`, which is the wrapper that drives the same chain. What you should see at the end is a report under assets/ named for the fund code; the README's worked example is fund-163417.html.

Claude Code users get a plugin path instead, with three slash commands:

/plugin marketplace add wbh604/fund-guy-skill
/plugin install fund-manager-alpha@fund-guy-skill

Those commands map to /analyze-fund 163417 for the full audit, /quick-check 163417 for a five-minute pass, and /find-similar 163417 for the holdings-overlap reverse lookup. Cursor, Windsurf, Codex, and Devin are handled by the repository-level AGENTS.md and the .agents/ and .cursor/ skill directories.

The 12-month verification window is also the main limitation

Every judgment in the report is anchored to price action 12 months after a trade. That is the design's strength and its ceiling. A position opened in the most recent quarter has no verdict yet, so the most current part of the book is structurally the least audited; the README's own framing of the method implies as much by tying outcomes to the following year. If you want to know whether the manager is currently positioned well, this tool cannot tell you, and it does not pretend to.

There is a second boundary. The score is a fixed weighted composite (timing 35%, control 35%, excess quality 30%), and the README says every point is recomputable, but recomputable is not the same as predictive. A manager who scores 62 has been shown to sell poorly over the past eight years; nothing in the README claims that predicts the next eight. The README also warns that the htmlpreview.github.io viewer often fails to run embedded scripts, which makes a working report look like a zero score and a blank chart. If you preview a gist-hosted copy in that viewer, you are looking at a broken renderer, not a broken analysis.

The report is also a single large HTML file. That is convenient for sharing and offline reading, and awkward for anything programmatic: there is no documented JSON output or API in the repository, so downstream tooling would have to parse the HTML or call the analysis scripts directly.

Alternatives and where fund-guy-skill differs

The obvious alternative is the platform score itself, and the README defines the project against it. A platform rating is a single number derived from returns and tenure, delivered as a service you cannot inspect. fund-guy-skill produces an audit trail instead: each trade matched to later price action, each conclusion traceable to a script, and unclear items explicitly marked unverified. The trade-off is that you get an artifact to read, not a feed to subscribe to, and you are responsible for running the scripts.

A second alternative is the manual approach: read quarterly reports yourself and compare holdings quarter over quarter. That gives you full control and no dependency on akshare or baostock, but it is exactly the work the README describes as unscalable. The skill's answer is the copy-trading index, which quantifies how much of the excess return survives waiting for the quarterly report. For the 163417 example the README reports that 80% of the excess remains, and that the entry-timing edge is worth only 4.4 points a year. That number is the strongest argument against the manual approach and against chasing the manager's entries, and it is also the kind of claim you should re-run for your own fund rather than accept from a demo page.

Licence, maintenance, and the cost of re-running

The repository is MIT licensed, which permits commercial use and modification provided the copyright notice and permission notice are retained. The README does not discuss attribution requirements beyond the standard licence text, and nothing in the repository addresses data redistribution rights for the underlying sources (akshare, baostock, Tiantian Fund), so if you plan to publish reports, that is a question for your own counsel rather than something the project answers.

On maintenance, the last push to the repository was on 2026-09-03, and the repository is not archived. There are no releases retrieved, so there is no versioned upgrade path to follow; the practical upgrade is a git pull plus a re-run of pip install -r requirements.txt. The deeper cost is data drift. The scripts depend on third-party endpoints, and the README already documents one fragile surface (the htmlpreview viewer). When an endpoint changes shape, the fetch scripts are what break, and you will notice it as a missing or stale intermediate file rather than a clear error in the report.

Editorial conclusion

Adopt fund-guy-skill if you already follow Chinese public funds and want a per-trade audit you can re-run from a fixed fund code, and if you are comfortable with an HTML report as the deliverable. Skip it if you need intraday or real-time signals, if you want a ranking service, or if your agent runtime is not one of the supported ones (OpenClaw, Claude Code, Codex, Cursor, Windsurf). Before trusting a report, verify the fetch scripts actually wrote fresh files for your fund code, confirm the score breakdown (timing 35%, control 35%, excess quality 30%) matches what the report shows, and check that any conclusion you plan to act on is not labelled as unverified.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. wbh604/fund-guy-skill on GitHub