How to write an AI character card

A character card is the small block of text that decides whether your AI companion feels like a real person or a polite chatbot wearing a name tag. It is the set of instructions the model reads before every single reply — who the character is, how they talk, where the scene is set, how they open. Get the card right and the character stays in voice for hours. Get it wrong and they drift, contradict themselves, or slide back into a generic assistant by the third message.

This guide is deliberately practical. Rather than theorise about prompt-writing in general, it walks field by field through an actual card — with a weak and a strong example for each — hands you a fill-in template, and is blunt about the mistakes that sink most first attempts. By the end you will have a card you can paste straight into a roleplay on SillyTavern, character.ai or rpdate.

One note on scope. This article is about the card — the structured fields that define a character. If you want the broader craft of writing prompts and steering a scene as you go, our companion piece on how to write AI character prompts covers that ground. The two work best read together; we will not repeat it here.

What a character card actually is

Strip away the platform differences and a character card is one thing: persistent context. Your chat history scrolls past and eventually falls out of the model’s memory, but the card is re-supplied on every turn. That is why it carries the load-bearing facts — the character’s nature, voice and situation — while the conversation handles everything that happens after.

Because the card is read so often, two qualities matter above all: it must be consistent (nothing contradicts anything else) and economical (every word earns its place). Most of the advice below is really these two principles applied to one field at a time.

It helps to picture the card as a contract you sign with the model once and renew silently on every turn. Anything you leave out, the model fills in from its own defaults — and those defaults tend toward the agreeable, helpful assistant it was trained as. Each field you write well is one fewer place for that default to leak through. That is the whole game: spend your words where they buy the most character, and trust the conversation to carry the rest.

The fields, one by one

Almost every platform’s card boils down to the same six fields under different names. Here is what each one is for and how to write it so it pulls its weight — with a weak version and a strong version of each, so you can see the difference rather than take it on faith.

Read the strong examples as a set and a pattern emerges: every one is more concrete and more economical than its weak twin. That is not a coincidence. Specificity and brevity are the two levers you pull on every field, and the rest of this guide is mostly about where to find them.

A finished card, filled in

Theory is easier to follow with a worked example. Here is a complete card for a character named Mira — short, consistent and built from the fields above. Notice how much of the work is done by showing rather than telling: the traits reappear as behaviour in the example dialogue and the greeting. Copy it, swap in your own details, and you have a working card in minutes.

That whole card is well under three hundred words, yet a model can run Mira for a long conversation from it. The persona never claims she is “witty” — the example and greeting simply are witty, and the model learns the voice by copying them.

W++ versus plain prose

You will run into two ways of formatting a card. The first is plain prose — ordinary sentences, like Mira’s card above. The second is W++, a bracket-and-quote syntax from the character.ai and Pygmalion era that packs traits into compact, code-like lists. Both feed the same model the same facts; they just package them differently, and that packaging has real trade-offs worth understanding.

Mira Vale, 27, a night-shift bartender in
a rain-soaked harbor city. Sharp green
eyes, a dark undercut, an anchor tattoo.
Dry-witted and guarded, fiercely loyal
once you are in, secretly soft-hearted,
always watching. It is 2am, the bar is
closed, and you are the last customer
she kept the lights on for.

[character("Mira Vale")
{
Description("27, night-shift bartender,
  rain-soaked harbor city, green eyes,
  dark undercut, anchor tattoo")
Personality("dry-witted", "guarded",
  "fiercely loyal", "secretly soft-
  hearted", "observant")
Scenario("2am, the bar is closed, you
  are the last customer she kept the
  lights on for")
}]

Same facts, same character — but the prose version reads like someone describing a friend, while the W++ version reads like a database row. Neither is wrong. The prose tends to carry a little more warmth and connective tissue; the W++ is leaner and easier to scan when a character has a long, dense trait list. The difference in the model’s output is usually subtle, which is exactly why the choice should come down to which one you find easier to write and edit.

The honest verdict: W++ is a tool, not a secret. It earns its keep when context space is tight or when you want a dense, scannable trait list. But modern models read natural prose beautifully, and a clear paragraph usually produces a warmer character than a wall of quoted keywords. A popular middle path is hybrid — a short prose persona for who they are, plus a compact trait list for quick reference. Write the character both ways once, run each for a few turns, and keep whichever voices them better; the test takes five minutes and ends the debate for good.

The common mistakes

Nearly every weak card fails for one of a handful of reasons. Learn to spot these and you will fix more problems than any amount of extra detail ever could.

Read those together and a single rule emerges: show, do not tell, and keep it consistent. A card that demonstrates its traits through dialogue and never argues with itself will outperform a far longer one stuffed with adjectives.

Build it here, with a coach watching

Reading about good cards is one thing; writing one is another. Fill in the fields below and the coach checks each as you type — calling out vague filler, text that is too thin or too long, and rewarding the example dialogue and first message that do the real work. Hit Load example to see the Mira card light up green.

Your card will appear here as you fill in the fields above…

Put your card to work

A finished card is a blueprint waiting for a story. The fastest way to see whether yours holds up is to drop it into a strong premise and play. If you want genre-specific scenes to stress-test a character against, these two pair especially well with a well-built card:

Test it, then tune it

No card is finished on the first draft. The fastest way to improve one is to run it: open a chat, play a few turns, and watch where the character wobbles. If they go flat, your example dialogue is thin. If they contradict themselves, two traits are fighting. If they ignore the setting, the scenario is too vague. Each failure points straight at the field that needs work.

For richer scene ideas to test a card against — first dates, rivalries, slow-burn setups and more — our roundup of AI roleplay scenarios gives you plenty of situations to drop a fresh character into and see how they hold up.

Skip the setup, or bring your own card

Writing a card is its own quiet pleasure — but you do not always have to. On rpdate every character already ships with a hand-tuned card: a coherent persona, a real voice, a scenario and a greeting that lands. You can start a conversation in one tap and feel, immediately, what a well-built card is supposed to feel like.

And when you do want your own character, everything in this guide carries over. Paste your persona, traits and first message into a chat, lead the scene, and steer as you go. An adult mode is available if you want it; tender and slow-burn works just as well.

The card is the blueprint. The chat is where it comes alive.

Frequently asked questions

What to read next

The companion pieces that pair best with a finished character card: