Create Your Own AI Girlfriend: Personality First, Looks Second
When you create your own AI girlfriend, the personality description decides whether you are still opening the app in week 6, and the appearance settings do not. A face is information the first 3 times you see it and wallpaper after that. Behavior is what you meet again in every message you send.
Key Takeaways:
- Appearance is a fixed asset. You consume it in the first session or two, after which it stops carrying new information no matter how good the rendering is
- Conversational behavior arrives fresh on every turn, which makes it the only part of a character that can still be interesting in week 6
- Most creation forms put the visual controls at the top and the personality field below them, which reverses the order that predicts whether the app stays open
- Experiments through the 1990s found people responding socially to machines that had no face at all, on the strength of wording and turn-taking alone (Reeves & Nass, 1996)
- A companion writes text that matches a description you wrote. The warmth in a good reply is assembled from your own instructions, and nothing on the other side is attached to you
The Face Runs Out of Information
An image stops carrying new information after a handful of viewings, and a description of how someone talks never does. That asymmetry is the whole argument. You picked the hair, the eyes, the outfit, and by the fourth session every bit of it has become the frame around the text box rather than something you look at.
Nothing is wrong with anyone for that. Attention drops off any stimulus that holds still, which is why the poster you loved in March is invisible by June and why the wallpaper on your phone is functionally a color. A companion’s picture has exactly one state. You used it up on day 1.
The text has a new state on every turn. By message 50 the character has either said something you did not see coming or she has not, and that single fact is doing most of the work of keeping the app open.
Horton and Wohl were describing radio and television audiences in 1956, and what they watched accumulate was a sense of knowing a performer’s manner: the turns of phrase, the reactions you could call in advance by the tenth broadcast (Horton & Wohl, 1956). Repeated behavior built the familiarity. The publicity photograph was never the part doing it.
Spend 20 minutes on the face and 40 seconds on the personality field, and you have put your time into the part with the shortest shelf life.
What You Re-encounter, and How Often
Rank the setup choices by how many times you will actually meet each one, and the ranking comes out backwards from the order the generator’s form presents them in.
| What you set | How often you meet it | What it decides |
|---|---|---|
| Face and body | 3 or 4 conscious looks, then it becomes furniture | Whether the app feels like yours in session 1 |
| Name | Every message, carrying no behavior with it | Almost nothing past the first hour |
| Voice, where there is one | Only the messages you play back | How much room the character takes up |
| Persona description | Every single reply, invisibly, because it is sent to the model each time | Manner, reply length, what she brings up unprompted |
| Saved memory of you | Every reply that reaches back past today | Whether week 6 is different from week 1 |
The persona description is the odd entry on that list, because you meet it constantly and never see it. It sits in front of every exchange as text the model reads before writing, so one lazy line gets re-executed 40 times a day. A vague description does not produce a vague character. It produces a generic one, which is worse, because generic is what the model falls back to, and that fallback sounds like customer support with a pet name. Sorting which settings on a custom ai girlfriend actually change her from the decorative ones saves most of the wasted effort.
Reeves and Nass spent the 1990s running experiments where people were polite to computers, took praise from them personally, and formed confident impressions of a machine’s personality from wording alone, none of it involving a face on the screen (Reeves & Nass, 1996). Whatever social response people had attached to how the thing behaved. Appearance was not the load-bearing part then, and adding a rendered portrait has not moved it since.
Treat the description field as something read aloud before every message you receive. Functionally, that is what it is.
Create Your Own AI Girlfriend by Writing Behavior, Not Adjectives
Adjectives compress to almost nothing. “Playful, caring, intelligent” fits half the characters ever written and hands the model close to zero constraint, while a line about behavior names what happens in a situation, and situations are what a conversation is made of. That is the core of how to create an ai girlfriend worth talking to: behavior over traits.
Five lines worth writing, in place of the trait list:
- Default reply length. “Two or three sentences unless I ask for more.” Nothing else changes the feel of an app that fast.
- What she does with questions. Say how often she asks one back. A character who asks something in 1 reply out of 3 reads as interested; one who asks in every reply reads like an intake form.
- What happens after silence. Write down what she does when you have not written for 2 days. Picks up mid-thought, asks where you went, or never mentions it at all. Three different characters. Leaving it unspecified is not a fourth, it is a default.
- One position she will not move off. Give her a stated opinion that contradicts one of yours. Cheapest available way to get a reply that is not a rewrite of your own message.
- One recurring detail. A city she claims, a job she complains about, a neighbor with a leaf blower. Something concrete enough to come back on its own.
Five behavioral lines beat 30 adjectives, and writing them takes less time than settling on an eye color.
Where Looks Do Matter
Appearance earns its 5 minutes for two reasons, and retention is not among them. A face fixes the voice you read the text in, and once that voice is set it is hard to overwrite. A face you find embarrassing is also a face you will not open with someone else in the room, and that tax gets paid even when nobody is watching.
There is a use where the whole argument collapses. Some people are there for the images and the conversation is the caption underneath. Different product, same interface, and if that is the use, the visual controls are the product and none of this applies.
Worth stating plainly, because the category invites the opposite reading: a carefully written persona produces text that fits a person, not a person. Whatever lands as warmth was assembled out of your own account of what warmth should sound like, handed back in a voice you also chose. Even a solid persona can shift under you when the platform swaps the model, the character you know answering differently after an update.
The Week 6 Test
One evening settles whether your character is doing any work at all. Open a fresh chat and send something with no emotional content in it: the thing that broke, what you ate, the errand you keep postponing. Read the reply, then ask whether those exact words could have come from any character in the catalog. If they could, the description is inert, and no amount of visual detail is going to cover for it.
Then run the strip test. Read the last exchange again with the picture out of your head, as though it had arrived in a plain gray window with no avatar attached. An exchange that still holds up is being carried by the character. One that goes dull was being carried by the image, and images stop carrying anything around week 2. Which is also why getting an ai girlfriend is a verdict to reserve for week three, not the first evening.
FAQ
Does an AI girlfriend’s appearance change how she talks? No. The picture comes from an image system that never reads your messages, so nothing about the face reaches the text. If a character sounds different after you change her look, the app edited part of her written description at the same time, which is a product decision rather than an effect of the image.
How long should the personality description be? 150 to 250 words of behavior beats 1,000 words of backstory. Long histories get summarized away or ignored, while behavioral instructions apply to the message being written right now. If you are past 300 words and none of it says what she does in a specific situation, cut it and start again from situations.
Can I change her personality later without starting over? Yes. The description and the saved memory are separate stores, so rewriting the description changes how she talks from the next message onward while what she knows about you stays where it is. Expect two or three strange exchanges where the old manner and the new instructions both show up in the same reply.
Why does she feel generic even though I filled in every field? Because trait words collapse into the model’s average character. “Kind, funny, a little shy” describes a default, and a default is what you get when the instruction constrains nothing specific. Swap one trait word for one sentence about behavior and the difference shows up in the very next reply.
The order the form puts things in is an argument about what is easy to render, not about what matters. Faces sell a signup. Behavior decides whether there is a week 6 for anything to be sold into. When someone builds a character and finds it has gone quiet on them a month later, the postmortem almost never turns up a problem with the eye color. It turns up 4 words in a personality box, doing the same nothing they were doing on day 1. So wherever you build one, Lona or anywhere else, put the effort into the sentences the character would say, not the face. The writing is what is still carrying the thing at week 6. The picture stopped mattering weeks before.
Sources
- Horton, D. & Wohl, R. R., “Mass Communication and Para-Social Interaction,” Psychiatry, 1956
- Reeves, B. & Nass, C., “The Media Equation,” 1996
