Daepak
Why generic AI answers fail on Korea
Ask ChatGPT "why did KOSPI fall?" and you get a paragraph about "global uncertainty." True, useless. The model has no live feed. It has no idea that on 2026-09-12 the index dropped 1.76% and lost the 7000 line, or that foreigners and institutions sold a combined 3.5 trillion won.
Korean market questions break in three specific ways. Ticker naming (005930, not "Samsung"). Data timing (Korea closes before the US opens). And units — won, not dollars.
Fix the inputs and the answers change.
What MCP actually does
MCP stands for Model Context Protocol. It's a standard way to plug live tools into an AI assistant. It's a USB port for data.
The ecosystem is already large. Directories list 9,800+ official and community MCP servers for Claude, Codex, Cursor and similar agents. Some pull prices. Some read filings. Some query databases.
You don't need to build one. You need to know what to connect.
For Korea, three connector types matter most:
| Connector | What it answers | Example source |
|---|---|---|
| Price/quote feed | What is it trading at now? | KRX, Upbit |
| Filing feed | What did the company actually report? | DART |
| Macro/calendar feed | What event is coming? | FOMC, BOJ, BOE calendars |
Without one, the model guesses.
Data-first prompting: the core habit
Data-first prompting means you paste the numbers before you ask the question. It reasons over evidence, not memory.
Bad prompt: "Is Samsung Electronics a good buy?"
Good prompt: "Samsung Electronics fell with the KOSPI's 1.76% drop on 2026-09-12. Foreigners and institutions net sold 3.5 trillion won market-wide. Korea's Sept 1-10 semiconductor exports rose 270.1% year on year and were 47.1% of total exports. What are the two strongest bull and bear arguments given only this?"
The second prompt has a testable answer. The first has a vibe.
Four rules for the data block:
- Give the date and time zone. Korea is KST, nine hours ahead of UTC.
- Give units. Won, dollars, percent — never bare numbers.
- Give the source. "KRX close" beats "I saw it somewhere."
- Ask for the counter-case. Force the model to argue both sides.
Korea-specific sources worth wiring in
DART is Korea's electronic disclosure system — the local equivalent of EDGAR. Numbers come from there, not news summaries. If you want revenue, margins, or share counts, that's the primary source.
KRX publishes index and stock data for KOSPI and KOSDAQ. KOSDAQ is the smaller, tech-heavy board.
Upbit is Korea's largest won-denominated crypto exchange. Its order books drive the "kimchi premium" — the gap between Korean and global crypto prices. The gap is a signal.
For macro, the calendar is the source. The week of 2026-09-14 to 09-18 packs FOMC, BOE and BOJ decisions together. Markets call it a super week. The dot plot and Powell's remarks are the two lines traders watch.
A prompt template you can reuse
Here's a structure that works across ChatGPT and Claude.
Context: [date, KST] [market] [what moved] [by how much]
Data: [numbers with units and source]
Question: [one specific decision or comparison]
Constraint: [answer in N bullets, cite which data point supports each]
Example, crypto version. Bitcoin traded at $77,129.04, down 0.232% over 24 hours, and US spot BTC ETFs saw roughly $450 million of net outflows for three straight sessions. Ethereum spot ETFs took inflows over the same stretch. Ask: "What does the flow divergence imply about relative positioning, and what would falsify that read?"
That's a question with a wrong answer available. Good.
Where the models still fail
Two failure modes survive good prompting.
First, stale training data. A model may "know" a rule that changed. Korea's crypto income tax is scheduled for January 2027. DAXA has asked for a review, citing gaps in cost-basis calculation and information exchange. A deferral bill to January 2030 was proposed. The government says it will proceed under current law. Ask a model cold and it may pick any of those three states.
Second, confident arithmetic. Always recompute percentages yourself.
The blog post that inspired this guide lists seven prompt uses: scenario building, chart reading, news impact, strategy backtests, portfolio risk checks, repeated-mistake review, and daily routine writing. Note what's missing. None of them predict price. The author's own framing: use AI to find what's missing from your judgment, not to tell you what to buy.
A short checklist
- Paste data before asking. Always.
- Name the source and the timestamp.
- Ask for the bear case explicitly.
- Cross-check any number the model computes.
- Re-ask after the close, not before.
If you want the numbers already assembled, the Daepak dashboard at /app keeps live pages for KOSPI names, won pairs and crypto. Copy a clean data block instead of hunting for it.
Related reading: how to vet AI stock analysis tools, and where Korean company numbers come from.
This is market education, not a recommendation to trade any asset.