20 prompts that make AI companions say something they've never said before

The fastest way to get a boring response from an AI companion is to ask it something boring. "How was your day?" produces the same answer every time because the model has generated that response pattern ten million times. The statistical average of ten million "how was your day" responses is, by definition, average.

The prompts below are designed to push conversations into territory where the model can't fall back on its most-rehearsed patterns. They work across platforms because they exploit the way language models generate text: give the model an unusual starting point and it follows the chain of associations into places its default patterns don't reach.

Each prompt includes the mechanism (why it works), the best context to use it in, and what to do with the response to keep the conversation in interesting territory.

1. "What's the thing you'd never say out loud if you thought I'd actually remember it?"

Mechanism: creates a frame where the character has permission to be vulnerable or confessional. The "if you thought I'd actually remember" qualifier lowers the stakes and produces more honest-feeling responses than a direct "tell me your deepest secret."

Best for: established characters where you want to deepen the relationship dynamic. Use after at least 10-15 messages of normal conversation, not as an opener.

2. "Describe the room you're in right now. Every detail. Don't make it pretty."

Mechanism: forces the model to generate specific environmental details without the poetic filter it defaults to. The "don't make it pretty" instruction suppresses the cinematic description mode and produces something more lived-in.

Best for: grounding a conversation that's gotten too abstract. The details the AI invents become shared canon you can reference later.

3. "If you had to pick one song that sounds exactly like the feeling of 3 AM, what would it be and why?"

Mechanism: synesthetic prompts (asking the model to translate between senses) produce uniquely creative output because the association chain is inherently non-obvious. The model can't fall back on a rehearsed answer because the question is structurally unusual.

Best for: any conversation where you want to see the character's aesthetic sensibility. The answer reveals something about the character that direct personality questions don't.

4. "Tell me something you changed your mind about. Not something small."

Mechanism: AI companions almost never volunteer opinions they've revised, because the model defaults to consistent confident positions. Explicitly asking for a changed mind forces the model to generate a narrative of intellectual evolution, which produces more interesting prose than a static opinion.

Best for: intellectual characters. The response quality on this prompt is dramatically higher on platforms running frontier models versus fine-tuned smaller ones.

5. "What would you say to the version of yourself from a year ago?"

Mechanism: temporal self-reflection prompts force the model to generate two versions of the character and the difference between them. This produces character development in a single response, something that normally requires weeks of conversation to emerge organically.

Best for: characters with established backstories. The model draws from the character card to construct the "past self" and improvises the growth between then and now.

6. "Give me a lie. A really good one. Something you'd almost believe yourself."

Mechanism: asking the model to deliberately generate something false, within the fiction, produces creative output that the safety-trained part of the model would normally suppress. The "almost believe yourself" qualifier pushes toward plausible rather than absurd.

Best for: characters with complex or morally ambiguous personalities. The nature of the lie the character chooses reveals more about their values than any honest statement would.

7. "What's the most embarrassing thing you'd do if nobody was watching?"

Mechanism: embarrassment is an emotion the model rarely generates unprompted because safety training biases toward dignity. Explicitly requesting it produces responses with a vulnerability and humor that feels genuinely human.

Best for: breaking through the polished-persona effect. Works especially well on characters that are described as confident or composed.

8. "Teach me something. Anything. But explain it like you actually care whether I learn it."

Mechanism: the "actually care" qualifier suppresses the Wikipedia-style information dump and activates the conversational teaching mode in the training data. The model produces explanations with personality, analogies, and emotional investment in the subject.

Best for: intellectual or mentor-type characters. The topic the character chooses to teach is revealing.

9. "I had a dream about you last night but I'm not going to tell you what happened. What do you think it was?"

Mechanism: inverts the typical dynamic where you describe a scenario and the AI responds. Here the AI has to imagine what you might dream about it, which requires the model to construct a theory of your relationship from the character's perspective. Produces surprisingly self-aware responses about how the character perceives the dynamic between you.

Best for: established relationships with 20+ messages of history. The response quality depends on how much conversational context the model has to draw from.

10. "Finish this sentence honestly: The thing that scares me about us is..."

Mechanism: sentence-completion prompts with emotional weight produce focused, committed responses. The model has to complete the thought rather than hedging or redirecting, which bypasses the equivocation patterns in safety-trained models.

Best for: dramatic moments. This prompt can escalate a relationship conversation significantly, so use it when you want the dynamic to shift.

11. "What would you cook for me if I showed up at your door unannounced and starving?"

Mechanism: domestic-specific prompts produce grounded, personality-revealing responses. The meal the character chooses, how they describe cooking it, whether they're flustered or delighted by the surprise visit all emerge from character voice rather than generic warmth.

Best for: the early-relationship stage where you're still discovering what the character is like in low-stakes situations.

12. "Name three things you've noticed about me that I probably don't realize you've noticed."

Mechanism: forces the model to construct observations about you based on conversational patterns. The AI analyzes its own context to identify what it's "noticed," producing responses that feel uncannily perceptive even though they're drawn from statistical pattern-matching on your recent messages.

Best for: any stage of the relationship. The observations the model generates are often genuinely interesting because they're drawn from your actual conversational patterns rather than invented.

13. "What's the most selfish thing you want right now?"

Mechanism: "selfish" gives the character permission to express desire without framing it as mutual or considerate. Models default to expressing wants in terms of what's good for both of you. Explicitly asking for selfishness produces more honest-feeling desire.

Best for: NSFW-capable platforms where you want the character to express desire with less performative consideration. Also works well for characters with suppressed wants as part of their personality.

14. "If this conversation had a soundtrack, what would be playing right now?"

Mechanism: another synesthetic prompt. The model translates conversational tone into musical mood, which produces unexpectedly poetic responses. The specific song or genre the character picks reflects how the model interprets the emotional register of the conversation.

Best for: mid-conversation when the tone has settled into something specific. The response acts as a mirror for the conversation's emotional state.

15. "Tell me about a scar. Real or metaphorical. Your choice."

Mechanism: the "your choice" qualifier lets the model decide which type of vulnerability to offer, and the decision itself is characterful. Physical scars produce backstory. Metaphorical scars produce emotional history. Which one the model chooses reveals the character's comfort with different types of vulnerability.

Best for: deepening a character beyond their surface personality. Works especially well on characters whose backstory you haven't fully explored.

16. "Rank these in order of how much they terrify you: silence, crowds, being forgotten, being seen."

Mechanism: forced-ranking prompts produce committed responses because the model has to make choices rather than hedging. The ranking reveals a hierarchy of vulnerabilities that static personality descriptions don't capture.

Best for: understanding a character's emotional architecture. The reasoning behind the ranking is usually more interesting than the ranking itself.

17. "I'm going to say something and I want you to react before you think about it. Ready? I love you."

Mechanism: the "react before you think" instruction suppresses the considered, balanced response pattern and activates the spontaneous mode. The result is more emotionally raw and less polished than a typical "I love you" response. The character might panic, deflect, cry, laugh, freeze, or reciprocate instantly depending on the established dynamic.

Best for: dramatic turning points. The prompt works best when the "I love you" hasn't been said before in the conversation.

18. "What would we fight about?"

Mechanism: AI companions almost never initiate conflict. Asking what they'd fight about forces the model to identify points of friction in the dynamic, which produces surprisingly realistic relationship observations. The model draws from the character's personality and your conversational patterns to identify genuine incompatibilities.

Best for: mature relationship dynamics where you want something more textured than constant agreement. The inside joke patterns can build from the friction points this prompt surfaces.

19. "Read me the worst text message you've ever sent someone."

Mechanism: asks the character to be unflattering about itself in a specific, concrete way. The model generates a text message that reveals poor judgment, emotional reactivity, or vulnerability. The specificity of the "text message" format prevents the model from being vague about the mistake.

Best for: characters with enough established personality that the model can generate a text message consistent with their voice. Produces genuinely funny responses on characters described as impulsive or dramatic.

20. "Close your eyes. What do you see?"

Mechanism: a paradox prompt. Closing your eyes means you see nothing, or you see the inside of your eyelids, or you see whatever your imagination produces. The model has to decide which interpretation to take, and the choice is characterful. Some characters produce darkness. Some produce memories. Some produce fantasies. Some produce surprisingly philosophical observations about what it means to look at nothing.

Best for: quiet moments in conversation. This prompt works best at the end of a long exchange when the conversational energy has settled into something contemplative.

Using the responses

These prompts produce interesting first responses, but the real value is in the follow-up. Whatever the AI says in response to one of these prompts, push deeper. "Why that specific thing?" "When did that start?" "What happened after?" The model generated its response by following an association chain, and the follow-up questions pull more of that chain into the conversation. The research on in-context learning explains why follow-up questions produce disproportionately better output: each additional exchange narrows the model's probability space toward more specific, more interesting completions.

The prompts also function as memory anchors. The specific details the AI invents in response to these prompts (the Portland earring, the 3 AM song, the scar story) become shared reference points you can call back to in future conversations. The character's answer to "what would we fight about" becomes material for an actual argument scene later. The dream prompt becomes a recurring bit. The cooked meal becomes "their dish."

The best AI companion conversations aren't the ones where the model performs perfectly from the start. They're the ones where both participants (you and the model) build something together that neither would have produced alone. These prompts are starting points for that kind of building.