"The Korean AI replies read okay but sometimes feel English-translated."
A symptom of LLMs trained heavily on English. This post breaks down chatroom Korean auto-reply tone into 4 areas + system-prompt patterns to lift operator voice from 80% → 95%.
Bottom line, 4 areas × 5 levels
| Area | Level 1 (awkward) | Level 5 (natural) |
|---|---|---|
| Honorific vs informal | Inconsistent | Auto-match per member |
| Address form | No address | "@OO-nim" / "OO-ya" natural |
| Shorthand / colloquial | Pure 안녕하세요 | "ㅋㅋ" / "ㅇㅇ" / "ㄱㅅ" natural |
| Sentence-ending variety | Always "~습니다" | " |
Guides from level 1 → 5 for each area.
Interactive tone-depth slider
Drag the slider to compare reply quality from level 1 → 5. Same member greeting, different tone depths.
Area 1, honorific vs informal auto-matching
Level 1 (awkward)
Persona system prompt has no consistency → same message returns mixed honorific / informal.
Member: 안녕하세요
AI: 안녕! 오늘 어때
Member greeted with honorific, AI replied informal = awkward + rude.
Level 3 (mid)
System prompt says "honorific consistent":
Member: 안녕하세요
AI: 안녕하세요, 오늘 어떠세요?
Honorific maintained, but identical tone to every member → creates distance with close members.
Level 5 (natural)
System prompt says "match per member". Honorific → honorific, informal → informal:
Member A: 안녕하세요 / AI: 안녕하세요, 오늘 어떠세요?
Member B: 안뇽 ㅎㅇ / AI: ㅎㅇ ㅋㅋ
Natural operator pattern in real chatrooms.
System prompt example
- When member uses honorific (~요 / ~습니다), respond honorific
- When member uses informal (~야 / ~지), respond informal
- Default to honorific with new members, match member tone once familiarity grows
Area 2, address form
Level 1 (awkward)
No address → unclear who the reply is for:
Member: 저 X 했어요
AI: 좋네요. 더 알려주시겠어요?
In a chatroom, "who's that for?" is ambiguous. Bot pattern.
Level 3 (mid)
Use @ mentions:
Member: 저 X 했어요
AI: @member 좋네요, 더 알려주시겠어요?
Address clear, but @ on every reply → sterile.
Level 5 (natural)
Mix address forms naturally:
Member 1 (close): 저 X 했어요
AI: 오 OO야 진짜?
Member 2 (formal): X 어떻게 해야 하나요?
AI: @OO-nim, X is like this...
Member 3 (question):
AI: 좋은 질문이네요. (no address)
Natural operator pattern.
System prompt example
- For close members (informal tone): "OO-ya" / just "OO"
- For polite members (honorific tone): "@OO-nim"
- For short replies or general questions, address optional
Area 3, shorthand / colloquial
Level 1 (awkward)
Perfect grammar, full honorific:
AI: 네, 그것은 좋은 의견이라고 생각합니다. 추가로 다음과 같은 점을 고려해주시면 좋을 것 같습니다.
Too formal for a natural chatroom. Strong bot signal.
Level 3 (mid)
Some shorthand:
AI: 네 그거 좋은 의견이에요. 추가로 이거도 고려해보세요.
Naturalness up, can go shorter depending on room vibe.
Level 5 (natural)
Mix shorthand / colloquial / short reactions:
Member: 와 진짜?
AI: ㄹㅇ ㅋㅋ
Member: 이거 어떻게 해요?
AI: 음 이렇게 해보세요
Member: 감사합니다!
AI: ㄱㅅㄱㅅ
Operator's natural chatroom tone.
System prompt example
- Short reactions naturally use "ㅋㅋ" / "ㅇㅇ" / "ㄱㅅ" / "ㄹㅇ"
- Reduce shorthand in info / deep replies, use precise vocabulary
- Calibrate shorthand frequency to room vibe
Pitfall
Excess shorthand → only "ㅋㅋ" / "ㅇㅇ" replies → obvious bot pattern. Balance shorthand with info replies.
Area 4, sentence-ending variety
Level 1 (awkward)
Every reply ends in "습니다" / "요":
AI: 그렇습니다. 좋은 의견입니다. 동의합니다.
Strong bot signal. Humans vary endings naturally.
Level 3 (mid)
Some variety:
AI: 그러네요. 좋은 의견이에요. 동의합니다.
Some variation, still short of natural.
Level 5 (natural)
Endings naturally varied:
AI: 음 그렇네 / 진짜? / 동의해요 / 그건 좀 / 흠 / ㄹㅇ
Different endings / interjections / natural fragments per reply.
System prompt example
- Distribute endings across ~요 / ~네 / ~지 / ~음 / ~군 / ~죠
- Use interjections (음 / 흠 / 오 / 아) naturally in short reactions
- Vary endings between consecutive replies in the same room
Path 80% → 95% operator voice
| Step | Task | Time |
|---|---|---|
| 1 | Pick 1 persona template | 5 min |
| 2 | Collect 50 of your chatroom messages | 10 min |
| 3 | Analyze your tone pattern (honorific ratio / shorthand / address) | 30 min |
| 4 | Explicitly state the 4 areas in system prompt | 1 hour |
| 5 | 10 sandbox test replies + rate | 1 hour |
| 6 | Retune prompt | 30 min |
| 7 | 1-week run + 👍/👎 rating | 1 week |
| 8 | Final retuning from feedback | 1 hour |
Total 56 hours + 1 week of operation = 80% → 95% voice match. The 5 principles in agent prompt writing guide.
Per-model Korean tone differences
User-perceived comparison (reference only). Actual experience varies with your room / members / machine.
The radar chart below maps four models across the four areas. Strengths and weaknesses jump out per model.
Deeper comparison in Qwen vs Gemma Korean comparison.
Frequently asked questions
Q. Won't loading all 4 areas into the system prompt overload replies?
No. System prompts guide tone, not length. Specifying the 4 areas raises naturalness.
Q. Can LLMs really auto-match per member?
Yes. LLMs look at the recent message context (member's last post) and match. "Match recent member tone" in the system prompt does it.
Q. If my room is casual, should shorthand / informal dominate?
Match the room. Social rooms 50%+ shorthand, info / business 10~20%. Room-type vibes in agent templates scenario matrix.
Q. "OO-ssi" vs "OO-nim"?
Subtle. "OO-nim" is more polite; "OO-ssi" has slight distance (Korean cultural nuance). Just "OO" for close members, "OO-nim" when politeness matters. "OO-ssi" rarely.
Q. What if shorthand like "ㄹㅇ" / "ㄱㅅ" feels off for polite members?
Per-member matching. No shorthand to polite-tone members, natural shorthand to close members. "Match member tone" in system prompt auto-handles it.
Q. Reviewers keep saying endings feel monotone, why?
Causes: (1) model limits (small models like Qwen 3B / Gemma 4B), (2) insufficient ending guidance in system prompt, (3) vague persona tone guide. Check (2) and (3) first. Use this post's example.
Q. How to make reply length feel natural?
Length is a separate variable. Short reactions (1 word ~ 1 sentence) dominate social rooms at 70%+, info rooms at 30~50%. Say "vary length by room mood / message type" in the system prompt.
Q. Can I train the LLM on my own replies?
Not directly (fine-tuning is separate). State your reply pattern explicitly in the system prompt (e.g. "operator uses short replies + shorthand + 'OO-ya' address"). Replyer's persona history tracks your tone evolution.
Next steps
To start auto-replies in your chatroom, download Replyer for your OS and follow the usage manual for the step-by-step guide.