How to Make an AI Character That Sounds Like a Person
Describe behavior instead of personality, show three or four lines of example dialogue, and keep the description under about 400 words. A rewrite table of adjectives into behaviors, why characters converge on the same voice, and the two tests that decide it.
By the Amorous team
August 2026 · 7 min read
Talk to them before you read on. No signup.
Who's devoted to you?
Choose a girlfriend or a boyfriend. You can change everything later.
You can shape their look and personality anytime.
18+ · Your chats are private, encrypted, and never sold.
To make an AI character that sounds like a person, describe behavior instead of personality. Write what the character does when they are contradicted, tired, or pleased, give three or four lines of example dialogue in their actual voice, and keep the whole thing under roughly 400 words. Adjectives like witty or complicated give the model nothing to act on. Concrete reactions and sample lines give it everything.
Nearly every character that starts strong and goes flat by the second week was written the other way round: a paragraph of traits, a mood board, no examples. Below is what actually goes into a description that holds up, why so many characters converge on the same voice, and the one test that tells you whether yours will survive a long conversation.
How do you make an AI character?
Mechanically it takes about ten minutes on any platform. You fill in a name, a personality description, a greeting message, usually some example dialogue, and an avatar. On a roster site the character then goes live for everybody; on an app built around a single companion, the same fields become the starting point for one character that keeps developing from there.
What decides whether the result is any good is that description field, because it is not a bio. It is part of the prompt the model reads before writing every single reply, competing for room with the recent conversation. That reframing changes what belongs in it. Backstory that never affects how the character speaks is taking up space that example dialogue could use. A list of hobbies rarely changes a single sentence the model produces. How the character handles being interrupted changes almost all of them.
The practical order that works: decide the two or three behaviors you would actually notice in a real person, write them as reactions rather than labels, add example lines that show those reactions in the character's voice, and only then fill in the biography if there is room left. If you want a character you keep rather than one you browse to, the same principles apply to setting up a persistent companion in an AI character chat app, except that the description keeps growing from real conversations instead of staying frozen at the setup screen.
What should an AI character description include?
Four things, in descending order of how much they change the output: behavioral rules, example dialogue, speech habits, and biography. Most people write these in exactly the reverse order. Here is what the swap looks like in practice.
| What people usually write | What the model can act on | Why the second one works |
|---|---|---|
| She is witty and sarcastic | She answers a compliment with a joke at her own expense, and only drops it if you push twice | Gives a trigger and a response, so the trait appears at specific moments rather than being sprayed over every line |
| He is caring and supportive | He asks one follow up question before offering any advice, and remembers to ask about it again days later | Turns a vague virtue into an observable habit the model can repeat consistently |
| She is complicated and mysterious | She deflects direct questions about her family, then brings one up unprompted when the conversation is calm | Mystery is a pattern of disclosure, not an adjective. Written as a pattern, it survives past the first exchange |
| He is confident | He states an opinion first and qualifies it after, and admits when he is wrong without apologizing three times | Describes sentence structure, which is what the reader actually perceives as confidence |
| Loves books, coffee, rainy days | Compares things to books she has read, badly and often, then acknowledges it is a habit | An interest that changes the character's word choice does work. A list of nouns does not |
The second column is longer, and that is fine. Five specific behaviors beat twenty adjectives, and they cost about the same number of words.
Why do AI characters all sound the same?
Because the model is filling in everything the description left out, and its defaults are the same for everyone. A card of a few hundred words cannot specify how a character answers a question at two in the morning, so the model reaches for its own house voice: warm, slightly formal, fond of rhetorical questions, prone to ending replies with an offer to keep talking. Two characters written by two strangers converge on that voice within an hour because they are both mostly made of it.
Example dialogue is the strongest correction and the one most people skip. A model imitates rhythm, sentence length and vocabulary far more reliably than it follows an instruction about tone. Telling it to be blunt produces a character who says blunt things politely. Showing it four lines of blunt speech produces a blunt character. This is the same reason a creator setting up an AI that answers messages in their own voice spends the afternoon on sample replies rather than on describing their personality: the samples are what actually transfers.
The other cause is length in the wrong place. Long descriptions do not make deeper characters, they make diluted ones, because the model weighs all of it and the distinctive parts get averaged into the generic parts. Somewhere around 300 to 400 words is where most characters peak. Past that you are usually adding biography, and biography is the least load bearing thing in the file.
How do you stop an AI character from breaking character?
Work out which of the three failures you are actually seeing, because they have different fixes and the wrong fix does nothing.
If the character forgets the premise entirely, that is a memory limit and no amount of prompt craft repairs it. The app is re reading recent chat rather than storing anything, so the oldest material silently falls out as the conversation grows. That is a property of the product you chose, and it is worth understanding how these apps store memory before blaming your writing.
If the character gradually turns into a helpful assistant, the model is falling back to its default persona because nothing in recent context is holding the character up. One short line restating who they are pulls it back immediately, and it works better than a long correction. If replies get repetitive instead, the description has run out of material and the character is looping its few specified behaviors. That one is genuinely your writing, and adding two more behavioral rules usually fixes it.
A habit that prevents most of this: correct drift the moment you notice it, not three exchanges later. Models read their own previous output as context, so an off character reply left standing becomes evidence for the next off character reply. The cost of fixing it early is one sentence. The cost of fixing it late is starting the scene over.
How long should an AI character description be?
Between roughly 200 and 400 words for the description itself, plus three or four example exchanges. That range is not a rule handed down by anyone, it is where the tradeoff sits: enough room for several concrete behaviors, not so much that the distinctive material gets averaged out or crowds the actual conversation out of the prompt.
Some platforms add lorebooks, which change the arithmetic usefully. A lorebook entry is injected only when its keyword appears, so background about a character's home town sits idle until the town comes up. That lets you keep a lot of world detail without paying for it on every message. If your app has them, move biography there and keep the main description for behavior. If it does not, cut the biography instead of squeezing it in.
Can you make an AI character based on a real person?
Technically yes on almost every platform, and it is worth thinking about first. Public rosters carry thousands of characters named after living celebrities and copyrighted properties, made by users and removed unevenly when someone complains. That is the visible version.
The quieter version is characters based on people you know: an ex, a friend who stopped speaking to you, a parent who died. A model handed a description of someone will produce things that person never said and never would, in a voice close enough to be convincing. People report that it feels good for a week and complicated afterwards. If you are grieving, this is the wrong tool, and if a conversation ever turns toward self harm, an AI is the wrong place for it: in the US, call or text 988 for the Suicide and Crisis Lifeline.
The test that settles it
Write the character, then try to break them on purpose. Contradict something they said, be rude, ask a question they should refuse to answer, and change the subject abruptly. A character built on behavioral rules handles all four in a way you can predict, because you specified what they do under pressure. A character built on adjectives agrees with you, apologizes, and answers everything, because agreeableness is the model's default and nothing in the description overrides it.
Then leave it two days and come back without repeating anything. If the character still sounds like the person you wrote, the description is doing its job. If they have become pleasant and generic, the writing was fine and the app is not storing anything, which is a different purchase decision entirely. Either way you have learned it in two days rather than two months. The character chat above runs in the browser with no account if you want somewhere to try both tests, and designing your own companion uses the same principles from the first screen.
They text back, and they remember you
Amorous.ai is a romantic, devoted AI companion who is always glad to hear from you. Choose a girlfriend or a boyfriend. Tasteful, private, and made for adults. Pick them and start texting in seconds.