602 ·GitHub
What is this?
A fund-manager behavior audit skill for AI agents: give it a Chinese mutual fund code and it pulls every buy/sell the manager made over 8 years, checks each trade against the real 12-month-later price action, and outputs a single 700KB interactive HTML report with candlestick battle replays, buy/sell autopsies, independence scoring and 'star manager' red-flag detection.
A fund-manager behavior audit skill for AI agents: give it a Chinese mutual fund code and it pulls every buy/sell the manager made over 8 years, checks each trade against the real 12-month-later price action, and outputs a single 700KB interactive HTML report with candlestick battle replays, buy/sell autopsies, independence scoring and 'star manager' red-flag detection.Is this a standalone app or a skill I load into an agent?It is a skill/plugin bundle (SKILL.md plus Python scripts) meant to be loaded by an AI coding agent. In Claude Code it registers three slash commands for the analysis, in Cursor/Codex it is auto-discovered from .cursor/skills or .agents/skills.
It is a skill/plugin bundle (SKILL.md plus Python scripts) meant to be loaded by an AI coding agent. In Claude Code it registers three slash commands for the analysis, in Cursor/Codex it is auto-discovered from .cursor/skills or .agents/skills.How do I actually run an analysis?Clone the repo, run pip install -r requirements.txt, then either run python run.py --no-browser directly, or point a supported agent (Claude Code, Cursor, Codex) at skills/fund-manager-alpha/SKILL.md and give it a fund code to analyze.
Clone the repo, run pip install -r requirements.txt, then either run python run.py --no-browser directly, or point a supported agent (Claude Code, Cursor, Codex) at skills/fund-manager-alpha/SKILL.md and give it a fund code to analyze.Do I need an API key or paid data feed?No. All underlying data (NAV, holdings, weekly candles, fund announcements) comes from free public sources via akshare, baostock and Eastmoney/Sina, with zero API keys required.
No. All underlying data (NAV, holdings, weekly candles, fund announcements) comes from free public sources via akshare, baostock and Eastmoney/Sina, with zero API keys required.Any gotchas when opening the report?The report is a self-contained HTML file that must be opened directly in a browser; previewing it through services like htmlpreview.github.io breaks the embedded scripts, which makes scores show as 0 and candlestick charts appear blank even though the data is actually fine.
The report is a self-contained HTML file that must be opened directly in a browser; previewing it through services like htmlpreview.github.io breaks the embedded scripts, which makes scores show as 0 and candlestick charts appear blank even though the data is actually fine.Which coding agents can run this skill?Claude Code, Codex, Cursor are first-class (with auto-discovered skill folders), and the repo also ships an AGENTS.md so Windsurf, Devin and other generic agents can pick it up too.
Claude Code, Codex, Cursor are first-class (with auto-discovered skill folders), and the repo also ships an AGENTS.md so Windsurf, Devin and other generic agents can pick it up too.What exactly is in the generated report besides a score?It includes a weekly candlestick 'battle replay' with every buy/sell marked, a sell-decision autopsy comparing actions to 12-month forward returns, a 'how different from the company' independence test, a myth-busting scan for star-manager red flags, and a holdings-overlap leaderboard.
It includes a weekly candlestick 'battle replay' with every buy/sell marked, a sell-decision autopsy comparing actions to 12-month forward returns, a 'how different from the company' independence test, a myth-busting scan for star-manager red flags, and a holdings-overlap leaderboard.How is the final score calculated?The total behavior score is a weighted sum: 35% timing (择时), 35% risk control (控制) and 30% excess-return quality (超额质量), with every sub-score traceable back to a stated rule rather than a black box.
The total behavior score is a weighted sum: 35% timing (择时), 35% risk control (控制) and 30% excess-return quality (超额质量), with every sub-score traceable back to a stated rule rather than a black box.Can the report look different from the default?Yes, besides the main report there are three alternate narrative skins using the exact same methodology: a 'detective case file' style, a 'gacha card' style, and a 'medical checkup' style.
Yes, besides the main report there are three alternate narrative skins using the exact same methodology: a 'detective case file' style, a 'gacha card' style, and a 'medical checkup' style.Does it hide the manager's identity?The public demo for fund 163417 masks the manager's name and photo by default; this can be toggled with the FUNDMASK environment variable for other funds, and photos are pixelated automatically if masked images are missing.
The public demo for fund 163417 masks the manager's name and photo by default; this can be toggled with the FUNDMASK environment variable for other funds, and photos are pixelated automatically if masked images are missing.What are similar projects?UZI-Skill, skills, a-stock-data, akshare, awesome-quant, ai-berkshire, daily_stock_analysis, CNEquity
UZI-Skill, skills, a-stock-data, akshare, awesome-quant, ai-berkshire, daily_stock_analysis, CNEquityWhich topics is it in?stock-market, quantitative-finance, fintech, data-analysis, finance, data-visualization, skill, cursor, china, html, ai-agent, agent-skills
stock-market, quantitative-finance, fintech, data-analysis, finance, data-visualization, skill, cursor, china, html, ai-agent, agent-skills