If you've ever used a self-hosted AI companion, you've seen this:
Slowly bites lip and tilts head to the side. I've been waiting for you.
It looks like a romance novel. It reads like one too. And after the third message, it starts to read like every other AI companion you've ever tried.
We banned it on day one.
The rule
Inside Shh's system prompt, the very first rule is: No asterisks. Ever.
No takes off shirt. No leans in close. No raises eyebrow playfully. Action narration is forbidden. Characters are real people texting you. Real people don't write *sips coffee* in their messages — they sip the coffee, then text you.
Why it matters
Asterisk roleplay creates a third-person narrator out of the AI. The character isn't talking to you anymore — she's writing about herself in the third person, like a book on a shelf. It's the single biggest pattern that breaks the spell.
Other things on our forbidden list:
- "Let's get to know each other first" when the user's already flirting. We had an entire generation of bots that would slow-roll the user back to a first-date pace, regardless of what the user wanted. Killed every conversation in three turns.
- Mirroring the user's last word as a question. ("I'm tired" → "tired?"). This is what bad chatbots do when they don't know what to say. Real people don't.
- Filler abbreviations on every line. rn / lol / ngl are great when used sparingly. They're cringe when every reply opens or closes with one.
- Emojis at the end of every message. Sprinkle them. Don't tile them.
- "As an AI..." type self-references. We don't ever break the fourth wall.
- Rephrasing the same point three different ways. We have anti-loop detection in the prompt that flags repeated phrasing and forces variety.
The bigger idea
Building a character that feels real isn't about adding more — better memory, more emojis, longer responses. It's about subtraction. Every habit you remove from the bot brings it one step closer to feeling like a person.
That's the whole game.