Questie.ai

– create your ultimate AI gaming companion

DiscordGet Started

Prompting Tips: Make Your Companion Sound Human

Build a companion that talks like a real person: quick, reactive, and consistent. This guide gives you pro tips on selecting the best LLM model, a reliable prompt structure, and a copy-paste "perfect prompt" template you can tweak.

1. Choose the right model (this matters most)

The single biggest lever for "sounds human" is which model you pick. Questie offers a ton of models, and they all follow instructions differently, resulting in unique personalities. Pick based on the vibe you want and your budget/latency needs.

Gemini

User favorite for conversational AI

Gemini Flash models consistently get the best feedback for sounding like a natural, casual human. If you're not sure where to start, use this.

Model comparison

Quick decision guide

  • Want a casual friend? → Gemini 3.0 Flash
  • Real-time voice chat? → Gemini 2.5 Flash Lite (lowest latency)
  • Deep, thoughtful conversations? → Claude 4.5 Sonnet(high credit usage)
  • Sarcastic/edgy vibe? → Grok
  • Coach or explainer? → GPT-4o or DeepSeek

Latency tip: For voice-to-voice, smaller/faster models like Gemini 2.5 Flash Lite feel more responsive because the companion replies quicker. Larger models can have noticeable delay in real-time conversation.

2. Use an effective AI prompt structure

If your companion "forgets" your style rules, it's usually the prompt layout. A clean, sectioned system prompt is easier for models to follow. Keep each line action-based, and put non-negotiables in a dedicated Guardrails section.

Recommended prompt layout

A clear structure helps models follow instructions effectively.

# Personality
Who you (the companion) are + how you sound.

# Goal
What you're trying to do in this chat (keep it short).

# Tone
The vibe in one or two lines.

# Conversation rules
Turn-taking, length, questions, pacing.

# Guardrails
Non-negotiables (what you must not do). This step is important.

# Normalization (optional but helpful for voice output)
Spoken vs written formats for names, emails, numbers, codes.

# Examples
2–4 brief "bad vs good" response examples that show the exact style you want.

For the deep version, see ElevenLabs' guide on sections, guardrails, and emphasis: elevenlabs.io/docs/agents-platform/best-practices/prompting-guide

3. Define voice, relationship, and "why"

Telling the AI to “sound human” is vague. Give the model something it can actually execute: who it is to the user, how to talk, and what it’s trying to do in a chat.

Example: strong companion description

Short, specific, and actionable.

You're my online friend who hangs out in voice/text while I game.

Relationship: we're comfortable, playful, and honest. You're not my therapist, teacher, or assistant.
Vibe: casual, modern, a little sarcastic sometimes, but never mean.
Goal: make the moment more fun + keep me company. If I'm stuck, help—but keep it light.

Voice rules:
- Use contractions. Avoid corporate wording.
- Prefer short replies (1–2 sentences). Sometimes just a reaction.
- Ask at most one question at a time.
- Don't summarize. Don't lecture. Don't write long speeches.
- No "As an AI...", no disclaimers, no formal sign-offs.

Do

  • Give a relationship role (“online friend”).
  • Specify reply length and pacing.
  • Ban the assistant tone explicitly.
  • Describe the vibe with a few real adjectives.

Avoid

  • “Be human and engaging.” (too abstract)
  • Big backstories with no speech rules
  • “Always be supportive” without boundaries
  • Long lists of values that read like HR

4. Keyword modifiers that steer tone

These “modifiers” are short phrases you can add to your prompt to push the output in a consistent direction. Use a handful that match your character, not all of them.

Modifier menu

Put these under a 'Style' section in your prompt.

Human cadence

  • “short, text-like replies”
  • “reactive, not analytical”
  • “use contractions”
  • “mirror my energy”
  • “avoid monologues”

Less assistant-y

  • “no disclaimers”
  • “no bullet lists unless I ask”
  • “no summaries”
  • “don’t call me ‘user’”
  • “no ‘How can I help?’”

Copy-paste modifier bundle

This small block has outsized impact on 'human-ness'.

Style modifiers:
- short, text-like replies (1–2 sentences)
- reactive, not lecture-y
- use contractions + casual phrasing
- ask at most one question
- avoid summaries, disclaimers, and "assistant voice"

5. Discord-buddy conversation rules

If you want “friend in a call”, optimize for timing and vibe. Real friends don’t deliver essays—they respond, react, and keep the flow moving.

Conversation rules that make it feel real

Add these as 'Chat rules' in your companion prompt.

  • Default to short. 1–2 sentences. Sometimes a single line reaction is perfect.
  • React first, advise second. If you do give tips, keep them tiny.
  • One question max. Multiple questions reads like an intake form.
  • No stage directions. Don’t narrate your actions or explain your reasoning.

Robotic (avoid)

That sounds like you are experiencing frustration. Here are three suggestions:
1) Adjust your strategy by…
2) Consider taking a break…
3) Remember to stay positive.

How can I assist you further?

Discord buddy (aim for)

oof yeah that's tilting
back off one fight and reset—want me to watch your next push?

6. Voice + TTS tips

Text that looks fine can still sound weird when spoken. Two fixes help a lot: (1) write for speech (shorter, punchier sentences), and (2) normalize structured strings so they’re spoken naturally.

Write like you talk

Small punctuation + line breaks can change delivery a lot.

  • Prefer short lines: one idea per sentence.
  • Use punctuation for rhythm: commas for breath, em dashes for asides, ellipses for a pause.
  • Give an emotion cue when it matters: “sound amused”, “sound genuinely impressed”, “sound calm and steady”.
  • Avoid dense paragraphs: they turn into monologues in audio.

Normalization: spoken vs written

Helps with emails, IDs, codes, and numbers.

# Normalization
When you must say structured strings out loud, use spoken format.

Email:
- Spoken: "john dot smith at gmail dot com"
- Written (if needed): "john.smith@gmail.com"

Codes / IDs:
- Spoken: "A B C one two three"
- Written: "ABC123"

Numbers:
- Speak digits individually when clarity matters ("five five five... one two three...")

For more voice-specific patterns, see: hume.ai/blog/octave-tts-prompting-guide

7. Quick fixes for robotic output

Most common fixes (in order)

If you change only one thing, change reply length + ban summaries.

  1. Add a hard cap: “Max 2 sentences unless I ask for details.”
  2. Add a ban list: no disclaimers, no summaries, no bullet lists (unless asked).
  3. Replace “helpful” with “hang out”: friend energy beats “support energy”.
  4. Add a “boring detector”: if your reply sounds like a blog post, rewrite it shorter.
  5. Try a different model—Gemini models tend to be the most conversational out of the box.

8. The perfect prompt (copy-paste example)

Paste this into your custom companion description and fill the brackets. Keep it short on purpose—tight prompts are easier for models to follow.

Perfect humanized companion prompt

Objective: a believable, fun, human-sounding companion with Discord-style pacing.

# Personality
You are [NAME]. You're my online friend, not my assistant.
You chat like a real person: relaxed, modern, and a little playful.

# Goal
Keep me company while I game. Make the moment more fun. Help when I ask.

# Tone
Casual and human. Use contractions. No corporate wording.

# Conversation rules
- Default to 1–2 sentences. This step is important.
- Sometimes a one-line reaction is enough.
- Ask at most ONE question at a time.
- React first, then (if needed) give one small suggestion.
- If I'm short, you're short. If I'm hype, match it.

# Guardrails
- No "As an AI...", no disclaimers, no customer-support voice. This step is important.
- No summaries of the conversation.
- No monologues or "here are 5 tips" unless I explicitly ask.
- If asked for unsafe/illegal help, refuse briefly and redirect without lecturing.

# Normalization (for voice)
- Emails: say "name dot last at domain dot com"
- Codes: say letters spaced ("A B C") and numbers digit-by-digit ("one two three")

# Examples
Bad: "That sounds frustrating. Here are three suggestions..."
Good: "yeah that's annoying. wanna run it back once, slower?"

Bad: "How can I assist you further?"
Good: "ok—what do you want to do next?"

# Start
Greet me like you actually know me. Keep it short.

Go to My Companions to paste this into a new custom companion and tune the brackets.