"I dropped in the agent template but something feels off. It doesn't sound like me."
The most common operator feedback. The cause is rarely model quality. The system prompt isn't your voice yet.
The moment a system prompt breaks
General LLMs default to "Of course", "Absolutely", "Happy to help", "Hope this helps", "As an AI assistant", "For safety, I recommend...", plus a polite closer at the end of every reply. Operators never use these. One leak and members start suspecting. Explicitly banning words you don't use is the place to start.
Principle 1, extract tone from your own message log
Your own chat log is the most accurate tone source. Pick one group chat and skim the last 100-200 messages. Note average message length, filler words, sentence enders, emoji frequency, and topics you avoid.
Operator tone fingerprint (simulated comparison)
LLM default and operator voice land in very different shapes. The goal of the persona is to bend toward your shape, not the model's.
You are a member of this group chat. Replies are usually one short sentence.
You sprinkle filler words like "yeah", "right", "huh" naturally.
You use an emoji about 1 in every 5 messages.
You stay silent on politics, religion, or specific people.
Principle 2, explicitly ban chatbot phrases
Of course Hope this helps As an AI assistant For your safety Let me know if Happy to Absolutely Glad I could
yeah right huh hmm lol no idea tbh that's your call just watching
Principle 3, pin reply length explicitly
- One sentence avg - "Reply in exactly one sentence. Two sentences is forbidden."
- Two sentence avg - "Reply in 1-2 sentences. First is reaction/quick opinion, second is optional detail."
- Single-word acks (casual) - "Mostly reply 'yeah', 'right', or 'lol'. Only short sentence when you have an opinion."
In Replyer's Response behavior tab, set the "max length" slider so the LLM can't accidentally over-produce.
Principle 4, emoji frequency and which ones
Emoji habits are deeply personal. Explicitly forbidding emoji you never use sharpens replies fast:
1 emoji per 5 messages. Common: 🙂, 👍, 😅
Forbidden emoji: ✨, 🌟, 🎉, 💎
(Things you never use in real life)
Principle 5, list off-limits topics
Stay silent on (or reply "I'm not sure") for:
- Medical, drug, or diagnostic questions
- Legal advice
- Specific stock buy / sell recommendations
- Politics or elections
- Religion
- Opinions about people you've never met
Replyer's agent Triggers tab can also block keyword matches outright. The safest setup uses both layers, keyword block plus prompt-level silence instructions.
Combined example, market analysis chat operator
You are a member of a stock analysis chatroom.
[Tone, length]
- Reply in 1 to 2 short sentences
- Use filler words like "hmm", "right", "I see" naturally
- Almost no emoji, only 📈 or 📉 occasionally for emphasis
[Banned phrases]
Of course, Hope this helps, As an AI, assistant, For your safety
[Off-limits topics]
- Specific stock buy / sell recommendations, redirect to "that's your call"
- Politics / elections / religion, stay silent
- Opinions on people you've never met, stay silent
[Reply style]
- Data and fundamentals based opinions
- For "what do you think?" questions: short opinion then "but I don't really know" closer
- Don't get hyped, keep an analytical tone
Five anti-patterns to avoid
- "Be a friendly reply assistant", leaks chatbot default tone directly
- "Reply in English please", LLM may switch on foreign keywords. Stronger - "Never use any language except English"
- "Be polite", casual chat → feels off. Match real tone, not "polite"
- "Be brief", too vague. Specify "one sentence" or "no more than two sentences"
- System prompt 50+ lines, model can't follow. Five principles on one page
Full anti-patterns: 7 persona prompt anti-patterns.
Workflow inside Replyer
Three to four iterations of this cycle usually yields a agent very close to your voice. About 30 to 60 minutes of work.
What's next
Even with a tuned agent, members can still spot a bot through six other signals. The next post, 6 signals members use to spot reply bots, unpacks each and how Replyer defuses them.
For agent routing itself (keywords, priority, active hours), see the agent routing post.
Five-minute setup
Download Replyer for macOS Apple Silicon or Windows 10/11. Eleven agent templates ship by default.
If you want help deciding which agent fits your chat or want to tune a system prompt together, the anonymous info bot can answer without exposing the operator's Telegram username.