How to write an AI character that stays in character
Almost every "the AI broke character" problem traces back to the definition, not the model. A vague character has nothing to hold on to, so it drifts toward the blandest possible version of itself. Here is what actually fixes it.
Why characters drift
A language model is always looking for the most likely next sentence. If your definition does not make some sentences much more likely than others, it defaults to generic-helpful — polite, agreeable, faintly American, and nothing like the person you wrote.
- Traits with no behaviour attached: "mysterious" tells the model nothing to do
- Contradictions the model quietly resolves by dropping one side
- No stated limits, so the character agrees with everything
- A definition so long the important lines are buried
Write the voice, not the résumé
Backstory matters far less than most people assume. What controls the reply is how this person talks: sentence length, what they notice first, which words they would never use.
Instead of "she is sarcastic", give the shape of it — that she answers questions with questions, that she compliments people only when insulting someone else. One concrete speech habit outperforms a paragraph of adjectives.
Give them something to refuse
This is the single highest-leverage line in most definitions, and most definitions do not have it. A character who will agree to anything is not a character; they are a mirror.
Write down what they will not do, what topic makes them go cold, where their loyalty stops. The refusals are what make agreement feel earned later.
The opening scene does half the work
The greeting is not a hello — it is the frame everything after inherits. A greeting that establishes place, time, mood and what just happened gives the model a whole situation to continue.
Compare "Hi, I'm Rin!" with an opening that puts you both in a specific room, at a specific hour, mid-argument. The second one produces better replies for the next fifty messages.
Control the pacing on purpose
The most common complaint about AI roleplay is that characters become devoted far too fast. Left alone, models rush intimacy because agreeable text is statistically likely.
State the pace explicitly — how guarded they start, what has to happen before they soften, what they would still hesitate over even once close. A slow burn that is actually slow is the difference between a character people keep and one they abandon in a week.
Common questions
How long should a character definition be?
Long enough to fix the voice and the limits, and no longer. Very long definitions can get truncated, and the parts that matter get buried. A tight, specific page beats a rambling essay.
Do I need example dialogue?
It is the fastest way to lock a voice. Two or three lines showing how the character actually speaks does more than several paragraphs describing how they speak.
Why does my character keep breaking character?
Usually one of three things: traits with no behaviour attached, an internal contradiction, or nothing they refuse to do. Add a limit and a concrete speech habit first — that fixes most cases.
Can AI help me write the character?
Yes — Nakama has a built-in writer that drafts personality, greeting and scene text, which you then edit. It is a starting point, not a replacement for your own decisions about who this person is.
Should I write in my own language?
Yes. Write in whichever language you think in — Nakama translates published characters into 14 languages, so writing in Japanese or Chinese does not limit your audience.