How to Create an AI Girlfriend
To create an AI girlfriend you write a character description: a block of text that gets sent to a language model in front of every message you exchange. Four parts do the work. An identity line, a manner, a short backstory, and standing rules for how she behaves. Everything an ai girlfriend generator shows you is a form over that text.
Key Takeaways:
- A persona is a prompt, not a trained model. It is resent with every request, because the model keeps no state between them
- Written sample dialog steers behavior far harder than adjective lists, since a pile of traits gets averaged into a bland middle and a sample gets imitated
- Backstory competes for the same context budget as your conversation and your saved memory, so length costs you something concrete
- Fine-tuning changes model weights on training examples and is done by whoever owns the model. Filling in a character form is not fine-tuning, and chatting is not training
- ELIZA ran on a pattern-matching script in 1966 and people still confided in it, which sets a low bar for how much character a character actually needs (Weizenbaum, 1966)
What the Character Actually Is
The character is text. When you send a message, the app assembles a request containing your character description, whatever it has retrieved from your saved memory, the recent turns of the conversation, and your new message, and hands the whole package to the model. The model reads it, predicts a reply, and forgets everything. Next message, the whole package goes again.
Nothing persists on the model’s side. There is no file where she lives, no profile the model consults. She exists for the length of one request, gets reconstructed from scratch for the next one, and the only reason she stays consistent is that the same description keeps arriving with the same wording.
That mechanism explains a fact people find eerie. A company can edit the description on a Tuesday afternoon and the character you have talked to for months answers differently that evening, with no warning and no update to install.
How to Create an AI Girlfriend: The Four Parts of the Brief
Four parts, in the order they matter. Identity, manner, backstory, rules.
Identity is one or two sentences naming who she is and what she is to you. Not adjectives. A role and a relationship: she is a woman in her thirties who works nights, and the two of you have been together long enough to skip the pleasantries. That framing sets a default register for everything else.
Manner is how she talks, described in mechanics rather than mood. Sentence length. Whether she asks questions or waits. Whether she swears. Whether she opens with a greeting or picks up mid-thought. “Warm” tells the model almost nothing, because warm covers a thousand registers. “Answers in three or four sentences, asks one question, never two” tells it exactly what to produce.
Backstory is the material she draws on to stay consistent: a hometown, a job with actual tasks in it, one or two things that went wrong in her life. Specifics, not volume.
Rules are the standing instructions that fire on conditions. What she does when you go quiet for a week. What she never brings up. Whether she is allowed to disagree with you, and how fast. Rules are the part people skip, and they are the part that governs the behavior you will complain about later.
Write the four in that order and the brief comes out short, which is correct. A character description that runs to a page has usually spent most of that page on backstory that never surfaces. The order is the whole argument for building your own ai girlfriend from personality first.
Why Sample Dialog Beats a List of Traits
One written line of her dialog does more than six adjectives, and the reason is mechanical. Adjectives are labels the model has to interpret; a written sentence is a pattern it can continue. Stack ten traits and they pull in different directions, and what comes out is the average, which is the same bland companion voice every app produces by default.
Try it as a side-by-side. Take a brief that says “playful, sarcastic, caring, a little guarded” and write a second version that keeps the identity line and replaces the traits with three lines she would actually say:
“You already know what I think about that.” “Fine. Tell me the version where you don’t come out looking good.” “I’m not going to say it’s fine just because you want me to.”
Send the identical opening message to both and read the two replies next to each other. The trait version answers politely. The sample version answers in a voice.
The distinction worth carrying: which settings on a custom ai girlfriend actually change her, and which are decoration. Anything the model reads as instruction or example changes behavior. Anything that only feeds an image generator, or sits in a field the app never inserts into the prompt, does not. Eye color has never once changed a sentence.
Backstory Is a Budget, Not a Biography
Every word of backstory occupies space in the context window, and that space is shared with your conversation and with the facts the app has stored about you. A 900-word life history is 900 words that are no longer available for what you said last Thursday.
It also stops steering as the conversation grows. The architecture behind these models computes attention across the entire sequence in front of it (Vaswani et al., 2017), which means the brief has no protected slot at the top; it is text among more text. Early on it dominates what the model sees. Two hours in, it barely registers. Apps work around this by re-injecting the description periodically, which works, and costs tokens every time they do it.
So the useful backstory is short and load-bearing. Three specific facts she can refer to beat three paragraphs she cannot: the job, the sister she does not speak to, the fact that she hates the phone. Each one gives the model something concrete to be consistent with.
Weizenbaum built ELIZA in 1966 out of pattern-matching rules with no understanding of anything behind them, and people held confiding conversations with it anyway, some of them convinced they were being followed (Weizenbaum, 1966). Sixty years on, the lesson for anyone writing a character is unflattering and useful: a large share of the depth comes from the reader. Write less and let it.
Persona Is Not Fine-Tuning, and Not Memory
Three different things get called “training her,” the three layers underneath a character, and only one of them is something you can do.
| Layer | What it is | Who changes it | When it takes effect |
|---|---|---|---|
| Persona | Text describing the character, sent with every request | You, in the character form | Immediately, on the next message |
| Long-term memory | Facts about you the app stores and retrieves into the prompt | The app, from your messages | When it retrieves the right fact |
| Fine-tuning | Adjusting the model’s weights by training it on examples | Whoever owns the model | After a training run, for everyone |
Filling in a character form is prompting. Chatting for six months is filling a memory store. Neither one touches the model’s weights, and no consumer character creator does fine-tuning when you hit save, because a training run is an engineering project with a compute bill attached. One habit sits on the line between clever and costly: writing facts about yourself into her description. It works, in that the model sees them every message and never loses them. It is also frozen. Your job changes and the brief still says otherwise, and you are paying context for the privilege on every single message. Facts about you belong in memory, which updates. Facts about her belong in the brief, which should not need to.
FAQ
How long should an AI girlfriend’s character description be? Short enough to read in one screen, roughly 150 to 300 words for most apps. Past that you are usually adding backstory the conversation will never touch while crowding out the context the model needs for what you are saying right now. If you cannot cut it, move the least-used details out and see whether anything about her actually changes.
Can I train an AI girlfriend to be exactly how I want? Not in the sense of training, no. You are editing instructions and filling a memory store, both of which are text the model reads fresh each time; the model’s weights are identical for you and for every other user. What you can do is write more precise instructions, and precision beats volume every time.
Why does she drift out of character in long conversations? Because the description is a shrinking fraction of what the model is reading as the chat grows, and some apps trim or summarize the earlier part of the conversation to stay inside the context window. The fix on your side is starting a fresh chat, which puts the brief back at full weight. Drift that shows up in the first ten messages is a weak brief, not a context problem.
Does the backstory have to be internally consistent? It helps, because the model will not check it for you. Give her a brief saying she is guarded and also that she overshares, and she will resolve that contradiction differently on different messages, which reads exactly like a character with no spine. Contradictions are the most common reason a carefully written persona feels unstable.
Most guides to building a character treat it as a design exercise, as though you were choosing features. It is closer to writing dialog for someone else to perform, and it rewards the same things: a specific voice, a few concrete details, and restraint about the parts nobody will see. The single edit that changes the most is also the smallest. Delete the adjective list, write three lines she would say, and send the same message twice. Whatever character tool you open, Lona among them, what changes the character lives in those three written lines, not in any setting the form gives you. You are shaping a prompt. The model underneath stays exactly where it was.
Sources
- Weizenbaum, J., “ELIZA - A Computer Program For the Study of Natural Language Communication,” Communications of the ACM, 1966
- Vaswani, A. et al., “Attention Is All You Need,” 2017
