tw-stock-etf-advisor

Taiwan Stock & ETF Advisor

This skill reproduces a research-to-tracking workflow for Taiwan equities. It has three actions. Figure out which one the user wants from their phrasing, then follow that section. The non-negotiable rules below apply to all three.

Core rules (why they matter)

1. Fetch real data — never guess, never derive prices. Stock prices and ETF weights change every day. A price computed from "holding value ÷ shares" is a stale close, not a quote. Always pull the live number with agent-browser. We learned this the hard way: a derived price was off by 6% the same morning.

2. For ETF holdings, prefer the fund company's official site over Yahoo. Yahoo only shows the top 10 holdings and its "資料時間" can lag by weeks (we saw 4/1 data on 6/1). The 投信 (fund company) sites publish the full constituent list at T+1. Use Yahoo only for live quotes, not for holdings. Exact URLs and extraction recipes are in references/data-sources.md.

3. Always state the data-as-of date and surface staleness. Each figure carries its own timestamp (holdings vs. weights vs. price). If something looks old, say so rather than presenting it as current.

4. This is not professional investment advice. End every recommendation with a short risk/disclaimer line. Taiwan AI-theme stocks are volatile with deep pullbacks — remind the user to verify the latest price/financials and honor stops.

5. Persist to Obsidian by delegating to Eliot. This skill does the analysis and data fetching; the vault writes (analysis note, [[stock]] project log, holdings ledger) go through the Eliot skill so they follow the user's templates and approval flow. See references/obsidian-tracking.md.

---

Action A — ETF common-holdings analysis & recommendation

Triggered by: "0050 跟 0052 共同持股", "推薦我買哪一檔", "哪一檔個股推薦購買". The user names two (or more) ETFs and wants overlapping holdings plus a reasoned pick.

1. Fetch full holdings for each ETF from the official 投信 site (not Yahoo). See references/data-sources.md for the URL per ETF and the exact agent-browser extraction recipe (the tables are div-based — read innerText around a known holding, don't rely on querySelectorAll('table')).

2. Compute the intersection by stock code. Build a table of common holdings with each ETF's weight side by side, sorted by weight. Note which large holdings are not shared (e.g. 0050's financials/traditionals that a pure-tech ETF lacks) — it explains the overlap's character.

3. Recommend stock(s) with reasoning. If the user asks for one, give one; if they ask for "at least N", give N and spread them across different links of the supply chain (e.g. material / thermal / system) so the picks aren't redundant. For each pick explain why (weight rank in both ETFs, structural theme, what it does) and briefly why not the obvious alternatives (e.g. don't single-bet 台積電 when the ETF already over-weights it).

4. If the user wants an entry/exit plan, fetch the live quote (Action C's price recipe) and give: reference price, buy zone, 停利 target (%), 停損 (%). Anchor percentages to the live price; flag any stock that already ran up hard today as "勿追高".

5. Offer to persist the analysis via Eliot (stock note + [[stock]] log). Don't write without the user's go-ahead.

---

Action B — Record a trade to the holdings ledger

Triggered by: "我買了 <股號/股名> <股數> @<價>", "賣出 <股號> 一半 @<價>". The user is reporting an actual transaction.

1. Parse: date (default today), action (買/賣), code, name, shares, fill price.
2. Pull 停損/停利 from the most recent analysis note for that stock if available; otherwise leave blank or ask.
3. Delegate the write to Eliot: append one transaction row and update the ## 持有中 position summary (cost average, total cost). Log a one-liner to [[stock]]. The exact ledger format is in references/obsidian-tracking.md.
4. If the fill price sits outside the analysis's suggested buy zone, mention it once (a neutral note, not a scold).

---

Action C — Check current holdings & suggest selling

Triggered by: "檢查我的持股", "我現在該賣什麼", "review my holdings", "should I sell anything". This is the action that closes the loop: read what the user owns, get fresh data, and produce a per-holding sell recommendation.

1. Read the ledger. Get the ## 持有中 positions from the holdings ledger (read is lightweight — obsidian read path="Eliot/Notes/2026/stock-holdings.md", or ask Eliot). For each position you need: code, name, shares, cost average, recorded 停損 and 停利 target.

2. Fetch fresh data per holding:

   • Live quote from Yahoo (price recipe in references/data-sources.md).
   • Current ETF membership & weight: re-fetch the relevant ETF holdings from the 官網 and check whether the stock is still a constituent and at what weight.
   • Technical indicators from Histock (recipe in references/data-sources.md): MA5/10/20, K9, D9, RSI6, RSI12, MACD. These are daily-close values (T-1 during market hours).

3. Evaluate technical signals per holding (supplement the fundamental sell signals):

   Signal	Condition	Interpretation
   均線空頭排列	price < MA5 < MA10	short-term downtrend, caution
   均線多頭排列	price > MA5 > MA10 > MA20	uptrend intact, support valid
   KD 死亡交叉	K < D and both declining	bearish momentum
   KD 黃金交叉	K > D and both rising	bullish momentum
   KD 超買	K9 > 80	overbought — consider scaling out
   KD 超賣	K9 < 20	oversold — possible bounce, but respect stops
   RSI 偏弱	RSI6 < 40	selling pressure dominant
   RSI 偏強	RSI6 > 70	buying pressure dominant
   MACD 正轉負	MACD crossing below 0	medium-term trend weakening

   Technical signals alone are not sell triggers — they add context to the fundamental sell signals below. E.g. "停損 approaching + KD death cross + RSI weak" strengthens the case for exit, while "near 停損 but KD golden cross" suggests watching one more session.

4. Evaluate all four fundamental sell signals per holding:

   Signal	Condition	Suggestion
   停損觸價	live ≤ recorded 停損	停損出場 — discipline, exit
   停利觸價	live ≥ 停利 target	分批停利 — scale out, let rest run if trend intact
   跌出 ETF 成分	no longer in 0050/0052 holdings	留意/考慮減碼 — weakened fundamentals signal
   權重明顯下降	ETF weight well below the level recorded at buy	留意 — losing index conviction
   報酬率檢視	report unrealized P/L % from cost regardless	context for the above

5. Present a holdings review table: code · name · cost · live price · 報酬率% · 技術面摘要 · 每個訊號狀態 · 建議 (續抱 / 分批停利 / 停損出場 / 留意減碼). Lead with anything actionable (停損/停利 hits) at the top. Include a per-holding technical summary row (MA position, KD state, RSI level) so the user sees the full picture.

6. Offer to update the ledger via Eliot if the user acts on a suggestion (record the sell, update 持有中). Don't write unprompted.

7. Close with the disclaimer line.

---

Mechanics & gotchas

• agent-browser: open → wait 3500 (a fixed wait — Yahoo's --load networkidle often hangs on ad/JS activity) → extract via eval --stdin heredoc → close when done. Full recipes and the exact selectors/text-anchors are in references/data-sources.md.
• Read that reference file at the start of Actions A and C — it has the per-site extraction code that is easy to get wrong (div-based tables, stale-data traps).
• Read references/obsidian-tracking.md before any persist step so the note and ledger formats match the user's existing vault conventions.