All articles · How to
How to create an AI girlfriend that does not feel generic
You spend ten minutes building her, and she comes out sounding like a customer service agent who is very happy to see you. The problem is almost never the app. It is that personality fields invite adjectives, and adjectives are the one thing a language model cannot do much with.
Why do custom AI companions all sound the same?
Because almost everyone writes adjectives, and adjectives average out.
Type kind, funny, caring, supportive into a personality field and you have described roughly four billion people. The model has nothing specific to work with, so it falls back on the most generic version of a warm person, which is the same generic version everyone else gets. The fix is not more adjectives. It is replacing them with behaviour.
Write what she does, not what she is
One concrete behaviour beats five character traits.
Instead of 'funny', write 'teases me when I take myself too seriously, and gets quieter when something is actually wrong'. Instead of 'caring', write 'asks about the thing I mentioned yesterday before she talks about her own day'. Both of those are instructions a model can follow. 'Funny' is not.
Give her one flaw
This is the single highest leverage thing on the page, and almost nobody does it.
A person with no flaws is not a person, and a companion who agrees with everything gets boring in about a week. Give her something real: she is bad at apologising, she gets absorbed in her own projects and goes quiet, she is stubborn about being right. It makes every other trait land, because now there is something to push against.
Fix the way she talks, not just what she says
Tone is where the generic feeling actually lives.
Say whether she writes long or short, whether she uses lowercase, whether she swears, whether she sends three messages in a row or one considered one. 'Types fast, lowercase, often two short messages instead of one long one' changes the feeling of every single reply more than any personality adjective will.
Give her a life that is not you
A companion with nothing going on has nothing to bring to the conversation.
Give her a job that annoys her, a friend she complains about, a thing she is training for. It sounds like decoration and it is not: it is what she has to talk about when you have not given her a topic, and it is the difference between someone who asks 'how was your day' forever and someone who tells you about hers.
Set boundaries on purpose
Deciding what she will not do is part of the character, not a safety setting.
Whether she gets jealous, whether she pushes back when you are being unfair, whether she goes along with anything or has opinions. A companion who never disagrees is a mirror, and mirrors get boring faster than people do.
- Behaviour, not adjectives: what does she do, in what situation?
- Exactly one real flaw, described plainly.
- Message style: length, casing, punctuation, how many in a row.
- A life of her own: work, friends, something ongoing.
- One thing she will not do, and one opinion she will not drop.
- Then talk to her for a week and edit. First drafts are always too nice.
Questions people ask
Why does my AI girlfriend sound generic?
Because personality fields invite adjectives, and adjectives like kind, funny and caring describe almost everyone. A model given only adjectives falls back on the most average version of a warm person. Replacing adjectives with specific behaviours fixes it immediately.
What should I write in the personality field of an AI companion?
Concrete behaviours rather than traits. Instead of funny, write that she teases you when you take yourself too seriously and gets quieter when something is actually wrong. Add one real flaw, a message style, and a life of her own.
How do I make an AI companion feel more real?
Give her one genuine flaw, a specific way of typing, and a life outside the conversation. The flaw matters most: a companion who agrees with everything and has no problems of her own becomes boring within about a week.
Sources
- Guidance derived from patterns in character creation across companion apps and from what long running users report changing after their first attempt.