Stopping AI from inventing fund manager opinions

What it does

This Agent Skill turns an LLM into a research assistant anchored entirely to Zheng Xi, an E Fund portfolio manager. It bundles his public statements from 2012–2026, a method framework distilled from those statements with original-text citations, and snapshots of his eight funds alongside a listing of roughly 27,000 market-wide funds. The skill answers questions, compares funds, and scores arbitrary funds against his framework, but every claim either points to a specific source document or is explicitly labeled as extrapolation.

The interesting bit

Most “AI investment analyst” projects are just prompt engineering that lets the model riff. This one treats hallucination as the primary enemy: if Zheng Xi never said it, the skill says so, and if it projects his method onto a new topic, it leads with a disclaimer. It is essentially a citation layer wrapped around a personality.

Key highlights

• Traceable Q&A with exact citations to interviews, reports, and manager notes
• Method framework backed by original quotes, not abstracted from third-party summaries
• Real portfolio data for his eight funds (current and former) for “words vs. deeds” checks
• Cross-market fund lookup and six-dimension scoring against his own criteria
• Packaged as a portable Agent Skill for Claude, WorkBuddy, and other platforms

Caveats

• Explicitly labeled as research-only and not investment advice
• Fund scores measure stylistic fit (“would he buy this?”), not absolute fund quality
• Live fund comparison requires Python scripts and network access; static corpus Q&A does not

Verdict

Worth a look if you are building retrieval-augmented agent skills for high-stakes domains where source fidelity matters. Skip it if you want a generic stock picker—this is deliberately narrow, opinionated, and anchored to one manager’s public record.