Korean
How to get natural Korean out of Syntic models, and what to check before you ship. The general rules for every language are on Work in other languages.
At a glance
| Language tags | ko-KR |
| Script and direction | Hangul, left to right, spaces between words |
| Token use | Measure with token counting on your own text; non-Latin scripts and long compounds usually take more tokens per sentence than English. |
Register and form of address
Speech levels matter. 합니다체 (formal polite) suits announcements and banking; 해요체 (polite, softer) suits most apps and chat support. Banmal (informal) is inappropriate for customers.
Numbers, dates and currency
Dates: 2026년 11월 5일. Currency ₩ without decimals. Large numbers often grouped by 만 (10⁴) in speech.
Common pitfalls
- Korean drops subjects freely; replies translated from English can sound stiff with repeated 당신 (you). Tell the model to avoid 당신 and use the customer’s name with 님, or omit the subject.
- Spacing rules (띄어쓰기) are strict in formal text; check output in evaluation.
A system prompt line that works
Writing the language instruction in Korean itself is a strong signal:
한국어로 해요체를 사용해 친절하고 간결하게 답변하세요. '당신'이라는 표현은 쓰지 마세요.(Answer in Korean in the polite 해요 style, kindly and briefly. Do not use 당신.)
Keep the rest of your instructions in whichever language your team maintains best; the model follows mixed-language system prompts reliably, and the reply language follows this line.
Test before launch
- Collect 30 or more real Korean messages, including short, misspelt and mixed-language ones.
- Have a fluent speaker grade replies for correctness, register and naturalness; then calibrate a rubric grader on their grades (Evaluate).
- Check rendering in your actual UI: line breaks, truncation and fonts.
- Re-run the set whenever you change the prompt or the model.