Zheng Xi Views Library
A source-traceable investment-research Skill built on every public statement by E Fund manager Zheng Xi — it quotes his actual words, reasons with his framework, and never fabricates.
SKILL.md mandates verbatim quotation, separates quotes from inference, forbids fabrication, and clearly disclaims investment advice; scripts run as single local commands with no hidden privilege expansion; dependencies limited to requests/bs4/lxml. Deducted: network-fetching scripts lack disclosed failure/rollback data-flow details; README suggests a broad Bash(python:*) allow rule; repo license metadata is NOASSERTION.
Instructions are self-consistent with well-scoped scripts, 7-day cache with --force refresh, synonym-search hints, and tenure-range filtering for former funds. Deducted: static review cannot confirm script error handling or failure feedback; no committed tests covering key paths.
Audience and scenarios are explicit (quote lookup, method, word-vs-holdings, scoring), triggers concrete, non-fit boundaries disclosed; Chinese corpus and mainland-accessible sources. Deducted: broad trigger surface ('any fund') risks over-triggering; live fetching unavailable in sandboxed environments (disclosed but still a limitation).
Well-layered docs, CHANGELOG with SemVer, MIT LICENSE, source disclosure, multi-platform guides. Deducted: skill.yml schema admittedly unverified against WorkBuddy; personal project with no explicit maintenance commitment or update guarantee.
README shows four realistic output examples in directly usable format and a one-command scoring entry, giving clear marginal value over impression-based answers. Deducted: static review cannot verify output correctness; examples are author-claimed 'real outputs' with unverifiable 2025–2026 performance figures.
Design is highly traceable: corpus_index., method.md claims each backed by quotes, fund data dated by quarter, sources named. Deducted: quoted corpus and data snapshots were not independently checked in this static review; static cap of 5 prevents claiming reproducibility.
- This is a static source review; no scripts were executed and corpus/performance figures in the shown examples (2025–2026) could not be verified.
- Fetch/score scripts depend on live access to external sites (e.g., Eastmoney); site changes may break them—mind cache freshness and verify key numbers.
- The README's suggested Bash(python:*) allow rule is over-broad; a narrower script-path whitelist is preferable.
- skill.yml's WorkBuddy schema is a best-guess by the author and may fail import without manual adjustment.
- The scorecard measures stylistic fit with Zheng Xi's approach, not fund quality; defensive/bond funds naturally score low and results must not be used as investment advice.
What does this skill do, and when should you use it?
This Agent Skill is grounded in the full corpus of public statements (2012–2026) by Zheng Xi, Deputy Head of Equity Investment at E Fund. Every answer traces back to a specific year and original text, backed by a distilled methodology where every rule cites his own words, enabling reasoned extrapolation on topics his corpus never covered. It ships with quarterly holdings and performance snapshots for all 8 of his funds plus a list of roughly 27,000 China mutual funds, supporting words-vs-holdings checks, peer comparisons, and a six-dimension 'Zheng Xi fit' fund score. It is strictly a research and learning aid, not investment advice.
Searches Zheng Xi's periodic reports, manager notes and interviews in references/corpus/ and returns quotes with sources; extrapolates on uncovered topics via references/method.md (first sentence declares it is not his own view); reads quarterly top-10 holdings and NAV/performance/scale snapshots for his 8 funds in references/fund_data/ for words-vs-actions checks; locates fund codes via fund_lookup.py across ~27,000 funds and fetches any fund's details online via fetch_any_fund.py into a cache; score_fund.py prepares data and computes mechanical metrics (concentration, turnover proxy, drawdown, etc.) for scoring against the six-dimension scorecard in references/scorecard.md.
- A tech-sector researcher verifying what Zheng Xi actually said about optical communications or AI compute, and when he turned bullish, with citations
- A fund analyst checking whether his stated bullishness on optical communications actually showed up in holdings, quarter by quarter
- A user applying Zheng Xi's framework to score a fund he doesn't manage (e.g. China Merchants Baijiu index fund) to ask 'would Zheng Xi buy this?'
- A content writer drafting a market outlook in the style of his quarterly reports, clearly labeled as a stylistic imitation
- A researcher comparing Zheng Xi's funds against any other China mutual fund (fetched online) on returns, drawdowns and holding style
What are this skill's strengths and limitations?
- Strong traceability: quotes must match the corpus verbatim, every methodology rule is backed by his own words, and the skill must distinguish 'his words / inference / needs verification'
- Real data: quarterly holdings and NAV/performance snapshots for all 8 of his funds ship locally, so words-vs-holdings checks don't rely on model memory
- Broader than one manager: includes a ~27,000-fund market-wide list for on-demand fetching, comparison and scoring
- Detailed cross-platform guidance: the README documents Claude Code, WorkBuddy, Cursor, ChatGPT, Gemini and ima deployment paths
- The score measures 'fit with Zheng Xi's style', not fund quality — defensive/dividend/bond funds score low by design and can be misread
- Live market-wide fetching depends on network access and the structure of public fund-data pages; there is no test suite, so page changes could break it
- Holdings are quarterly disclosure snapshots, not real-time; data for formerly managed funds spans each fund's full history and must be filtered to his tenure
- Highly vertical scope (Zheng Xi and Chinese mutual funds only); GitHub license is NOASSERTION while the README claims MIT — an inconsistency
How do you install this skill?
Native Claude Code install — macOS/Linux: mkdir -p "$HOME/.claude/skills/zhengxi-views" && cp -R SKILL.md README.md references scripts "$HOME/.claude/skills/zhengxi-views"/; Windows PowerShell: use the equivalent New-Item + Copy-Item commands from the README. Fully restart Claude Code afterwards. For Tencent WorkBuddy, drop the whole folder into its skills/ directory (a skill.yml manifest is included). For online fetching/scoring, run pip install -r requirements.txt (requests/beautifulsoup4/lxml). Deployment paths for Cursor, ChatGPT GPTs, Gemini and Tencent ima are documented in the README, though sandboxed platforms without network cannot do live fetching.
How do you use this skill?
After restart, just ask naturally, e.g.: 'What does Zheng Xi think of optical communications? When did he turn bullish?', 'He said he likes optical comms — do his holdings confirm it?', 'Score China Merchants Baijiu by Zheng Xi's standards — would he buy it?', 'Compare Zheng Xi's fund with Ge Lan's CMS Medical'. For topics outside his corpus, the skill falls back to the method.md framework and bold-declares in the first sentence that it is not his own opinion.
How does this skill compare with similar options?
The README explicitly distinguishes this skill from generic 'investment framework/methodology' skills: those abstract methods without citing source text and output stock picks, whereas this skill anchors everything in Zheng Xi's verbatim corpus, producing sourced quotes and verifiable forward reasoning, with 'no fabrication' as its core constraint.