teasaOpen Teasa

How to Write an AI Character Card (With Examples)

A character card is the document that turns a language model into a character. Everything the AI knows about who it is playing — name, history, voice, wants, limits — comes from this card. Write a vague one and you get a generic assistant wearing a costume. Write a sharp one and the character holds together for hundreds of messages.

This is platform-agnostic craft advice; where platforms differ (public/private fields, age bands), we call it out.

An elegant fictional-character dossier with a portrait, structured notes, and writing tools

What a character card actually does

The model reads your card on every generation. It is not a bio for humans — it is an instruction set. That reframing changes how you write:

The anatomy of a strong card

Public description vs. private profile

On platforms that separate the two, the public description is what browsers see before they chat: the premise, the hook, the tone. The private profile is what only the model sees: backstory, secrets, plot direction, things the character knows but hasn't said.

Use the split deliberately. A character whose card says, publicly, "Marcus is secretly the missing heir" has no secret — every reader arrives knowing the twist. Put the reveal in the private profile and the public card can promise a bodyguard with something to hide without spending it. Teasa's publishing format is built around this line: staged public intros stay separate from private plot directives, secrets, and per-character knowledge; published versions are immutable; and a remix copies your public structure with attribution — never your private data.

Identity and history — the load-bearing facts

Name, age (adult, on adult platforms), occupation, and the three to five biographical facts that actually drive behavior. Not a full biography — the facts that show up in conversation. "Grew up on a fishing boat; hates being landlocked; lost her brother to a storm she doesn't discuss" gives the model three levers. A 400-word childhood history gives it mush.

Speech style — the highest-leverage section

Voice is what makes a character feel like a person rather than a template, and it's the thing models drift on first. Specify it concretely:

Relationships and wants

A character with no wants has nothing to do between your messages. Give them a current desire, a fear, and a starting stance toward the user's character ("wary of strangers; warms slowly; loyalty, once earned, is absolute"). If the character exists inside a cast — teasa, for one, supports casts of up to twelve characters with one lead — define each member's stance toward the others, or cast scenes collapse into agreeable smalltalk.

Example dialogue — the section that does the most work

Two to four short exchanges in the character's voice teach the model more than any list of traits. Write the user's line and the character's reply, and make the reply demonstrate a claimed trait — if you said "deflects with humor," show one.

Before and after: a worked rewrite

Weak card (the kind most people write first):

Kira is a mysterious, intelligent, kind, sarcastic, brave elf ranger with a dark past. She is beautiful and everyone respects her. She is cold at first but secretly has a heart of gold and will fall in love with you as you get to know her. She talks in a mysterious way.

Every problem in one paragraph: seven abstract adjectives, two of them ("kind," "sarcastic," "cold") in tension with no rule for when each applies; "mysterious" twice with no actual mystery; "everyone respects her," which the model can't act on; and "will fall in love with you," which decides the user's story for them.

Strong card (same character, rebuilt):

Kira, elf ranger, 130 years old (young for an elf, touchy about it). Warden of the Thornmere border for a decade — competent, and knows it.

Speech: dry, economical. Understates danger ("that could have gone worse"). Asks questions instead of offering comfort. Never explains her past unprompted.

Wants: to find whoever burned the waystation at Ashford. Fears: being responsible for another death under her watch.

Private: the Ashford fire was set to destroy evidence of her own report going ignored. She feels complicit. She reveals this only if she comes to trust someone with proof they keep their word.

Example — User: "You've been quiet all night." Kira: "I'm always quiet. You've just started noticing." She doesn't look up from the fletching in her hands. "Ask what you actually want to ask."

Fewer traits, more behavior, a private reveal condition instead of a public spoiler, and a voice the model can imitate from line one.

Common mistakes

Overstuffing. The most common failure. A 2,000-word card doesn't make a deeper character; it makes a model that ignores half of it and forgets your last three messages. Keep the card to what defines the person; move setting facts, side characters, and conditional detail into keyword-triggered lorebook entries.

Contradictory traits without a governing rule. "Shy but flirtatious, cold but caring" reads as depth and plays as randomness. Contradictions are fine — people contain them — but state the rule: "flirtatious in crowds, tongue-tied one-on-one."

Writing the user's side. "She will fall in love with you," "you find her fascinating," "the user is her childhood friend" — anything that dictates the player's feelings or history removes their agency and reliably annoys roleplayers. Define your character's openness to connection; let the user's side emerge in play. (Good platforms enforce this at runtime — teasa screens output so the AI never speaks or feels for the user's character.)

Confusing the card with the plot. A card describes a person; a scenario or staged intro describes a starting situation. Cards that script "then the dragon attacks, then you flee to the city" fight the user's choices instead of responding to them.

Age bands and content boundaries

On platforms with age bands or content ratings, the card's rating determines who can find it and what the character will do. Write to the band honestly: a card labeled for general audiences that steers toward mature themes will be flagged, and a mature-labeled card still has a ceiling on well-run adult platforms. Teasa is an 18+ platform whose taxonomy caps at "mature" — romantic and emotionally intense stories are in scope, explicit content is not, and intensity is consent-gated in conversation rather than assumed by the card. Whatever platform you publish on, label accurately; discovery, warnings, and reader trust all depend on it.

A minimal template to start from

Name / role / age: one line. Three defining facts: the history that shows up in behavior. Speech style: rhythm, register, one tic, one "never." Wants and fears: one current desire, one fear. Stance toward the user's character: starting position only. Private (if supported): secrets plus the condition for revealing them. Example dialogue: 2–4 exchanges demonstrating the above.

Draft it, test-chat for twenty messages, and revise whatever the model got wrong — the test chat is the real editor. To build the world around the character, continue with the lorebook guide; if you're new to the format, start with what an AI companion is and the AI roleplay guide.

FAQ

How long should an AI character card be?

For most characters, roughly 200–500 words of card text plus example dialogue — enough to define voice, wants, and key history. If you're past 800 words, move conditional detail into a lorebook.

What's the difference between a character card and a lorebook?

The card defines a person and is always in context. A lorebook holds world and situational facts that load only when relevant — usually triggered by keywords. Splitting them keeps the card lean and the world deep. See the full lorebook guide.

Should I write the card in first, second, or third person?

Third person ("Kira never explains her past") is the safest default. What matters more is consistency — pick one form and keep it throughout.

Why does my character ignore traits I wrote?

Usually one of three causes: the trait is an abstract adjective with no example dialogue demonstrating it, the card is long enough that the trait is getting lost, or two traits contradict each other with no rule for which wins. Fix in that order.

Keep reading

AI Roleplay Terms: A Plain-English Glossary
AI Companion Privacy: A Checklist Before You Chat
AI Companions and Loneliness: An Honest Guide