"I spent days on the persona prompt but auto-reply feels off. What went wrong?"
Most common operator question. Seven anti-patterns plus symptoms / causes / fixes, illustrated with side-by-side examples and an impact radar so you can see which one is hurting you most. The inverse of the persona prompt writing guide.
How each anti-pattern affects auto-reply quality (simulated)
Axes are the dimensions Replyer's Diagnostics 6 heuristics (tone drift / duplicates / length anomalies) can surface. Anti-pattern accumulation collapses the shape inward.
Anti-pattern 1, prompt too long
You are friendly and professional yet sometimes humorous, maintaining politeness, answering in English but using some Korean words when natural, shifting between formal and casual depending on context... (2400+ chars continues)
You are a market-analysis peer.
- 1~2 sentences per reply
- Filler: "hmm", "right", "lol"
- Emoji only 📈 / 📉 for emphasis
- Forbid formal endings ("입니다")
- Say "not sure" when unknown
[5 few-shot examples]
Symptom Persona system prompt 2000+ chars. LLM forgets some rules or wobbles between contradictory directives.
Cause Operator tries to cram every situation / tone guide into the prompt. Persona becomes a "manual" instead of a single character.
Fix Keep prompt in the 700-1500 char range. 5-10 core tone guides + 3-5 few-shot examples. If you need more guidance, split into separate personas. See persona aging for rewrite procedure.
Anti-pattern 2, contradictory guides
Symptom Auto-reply lacks consistency. Some messages formal, others casual. The prompt contradicts itself.
Cause Both "formal tone" and "casual tone" written into the prompt. LLM picks different guides per message. Or few-shot examples drift from the stated guide.
Fix One direction for tone guidance. Either formal or casual, clearly. If ambiguous, set priority - "default formal, match casual if member opens casual". Few-shot examples must align with the prompt's guide.
Anti-pattern 3, abstract tone descriptions
Operator hasn't articulated their voice precisely. Persona gets only "vibes".
friendly but professional just the right energy not too heavy appropriately formal
friendly → endings -네 / -지 freely,
acks ㅋㅋ / ㅎㅎ often
professional → 1~2 info points/msg,
cite source when quoting
energetic → 1 emoji per msg,
response markers (ㄹㅇ / 맞)
Tone guides should be directly-executable behaviors for the LLM.
Anti-pattern 4, missing few-shot examples
Symptom Persona prompt has tone guides but few or no actual response examples. LLM interprets tone freely, drifting from operator intent.
Fix 3-10 few-shot examples. One operator response example per message type (question / acknowledgement / info / joke). Even if examples take up half the prompt, that's fine.
e.g. few-shot inside persona prompt:
[member] How do you see stocks today?
[you] oh today's US close is decisive lol just watching for now
[member] right lol
[you] ㄹㅇ jumping in there is just begging for stop-loss
[member] What do you think about this?
[you] hmm chart alone says buy signal but the macro is weak
See persona prompt writing guide.
Anti-pattern 5, no forbidden phrases
Auto-reply uses words / endings the operator never uses (e.g. English acronyms / formal endings / AI self-intros). Members suspect a bot.
Persona's forbidden_topics / hard-banned phrase fields are empty, so LLM defaults to its base training tone. List forbidden phrases:
- AI self-intros ("AI", "bot", "tool", "GPT" etc.)
- English acronyms the operator avoids ("BTW", "FYI" etc.)
- Formal endings ("입니다", "였습니다") for casual personas only
- Chatroom policy violations (ads / profanity / politics)
See responding when AI replies get caught.
Anti-pattern 6, operator-voice vs prompt mismatch
Prompt reads natural but auto-reply diverges from operator voice. Members read "operator tone changed".
Persona wasn't built from operator-voice analysis. Generic / abstract tone. Operator's own response data ignored.
Before writing the persona, collect 30-50 of the operator's own responses, analyze, then encode in the persona. Compare persona responses vs operator responses monthly (Sandbox). See persona A/B testing.
Anti-pattern 7, no per-chatroom persona separation
Operator runs multiple chatrooms with the same persona. Different room moods / topics get the same tone. Naturalness drops in some rooms.
- Info chatroom - formal + info-heavy persona
- Social chatroom - casual + acknowledgement-heavy persona
- Paid chatroom - deep response + 1:1 persona
Operator voice is the shared base + per-chatroom micro-tuning. See cross-chatroom info sync.
Self-diagnosis
FAQ
Q. Short prompt - won't auto-reply be thin?
LLM's base training + few-shot + operator voice align cover what a short prompt loses. A 1000-char prompt often beats a 2000-char one. Key: clarity + consistency.
Q. How many few-shot examples is enough?
5-10. More than that bloats the prompt. Under 5 lets the LLM interpret tone too freely.
Q. What if my forbidden-phrase list grows too big?
50+ forbidden phrases makes normal generation hard for the LLM. Cap at 10-20 core items. AI self-intros / words operator avoids / chatroom policy violations only.
Q. How often to rewrite the persona when operator voice shifts?
3-6 months. Operator's tone / vocabulary / register drifts naturally, persona prompt must follow. See persona aging.
Q. With 5+ personas across chatrooms, ops overhead?
Use persona zip to maintain a base + micro-tune copies. Ops overhead stays low. See moving Replyer to another PC.
Q. After finding anti-patterns, fix immediately vs gradually?
Major changes (full rewrite) get noticed by members as "tone shift". Use the 4-step gradual transition in persona aging.
Next step
Grab the build for your OS from the Replyer download page and follow the usage manual for step-by-step setup.