Character Card 101: The Hidden File That Makes Your AI Companion Feel Real

A character card is a JSON file with a defined schema. The most widely-used format originated with TavernAI in 2022 and was extended by SillyTavern (the most popular open-source frontend for character chat) into what's now known as the V2 character card spec. Most consumer apps including Janitor AI either use this format directly or accept it via import.

The core fields are straightforward in name but layered in practice. 'Name' is the character's name. 'Description' is a paragraph or two of who they are physically and personality-wise. 'Personality' is a more structured trait list. 'Scenario' sets the situation the chat begins in. 'First message' is what the character says before the user types anything — this is shockingly important, since it sets tone and grants the AI immediate momentum. 'Example dialogue' is 2-5 paired exchanges showing how the character should sound in conversation, which the model uses to calibrate its voice.

Advanced fields include 'system prompt' (instructions to the model about how to play the character regardless of conversation), 'world info' (a database of lore facts the character knows), 'tags' (for discoverability), and creator metadata. A well-engineered card uses all of these. A weak card uses two or three.

When you tell a generic LLM 'you are a moody anime girl named Yuna who likes coffee,' you get a generic LLM doing a moody anime girl impression for about three exchanges before it drifts back to being itself. The reason is simple: the model has billions of parameters trained on the entire internet, and a 200-character prompt is a feather trying to redirect a freight train.

A 50KB character card is a different kind of intervention. It includes 5-10 paired example exchanges that show the model exactly how Yuna sentences are constructed (clipped, deflective, occasionally sharp). It includes a personality matrix that says 'Yuna avoids direct emotional expression but reveals herself through rituals — ordering specific drinks, fidgeting with sleeves.' It includes a scenario that places the chat in a context Yuna's behavior makes sense in. It includes a first message that puts Yuna in motion before the user can drift the tone.

Result: the model has dozens of anchors pulling it toward consistent character behavior, not one weak directive. The character holds across 100+ exchanges instead of 3-5.

The character card concept emerged in the open-source AI roleplay community starting around 2021-2022. Pygmalion AI, an early NSFW-friendly model fine-tuned for roleplay, popularized the idea that you could ship 'a character' as a portable file that worked across different inference backends. TavernAI, originally a fork of KoboldAI's UI focused on roleplay, formalized the JSON schema that became the de facto standard.

When SillyTavern emerged as TavernAI's more actively-developed successor, it extended the format into what's now called the V2 spec, adding world info, alternate greetings, and more sophisticated metadata. By 2024 the format was so dominant that consumer apps building on top of LLMs started supporting it as an import format.

In 2026 the V2 character card format is effectively a standard — meaning a card you write today can be imported into Janitor AI, SillyTavern, Risuai, AgnAI, and dozens of other apps. It's one of the rare cases of a community-driven format outlasting the products that birthed it.

Reading enough top-rated cards on platforms like Janitor AI, you start to notice patterns in what works. The first signature is investment: great cards tend to be 8-15KB minimum, with 5+ example dialogues. Lazy cards are <2KB and have one example. The model can tell.

The second signature is specificity. Weak cards say 'she's confident.' Strong cards say 'she enters rooms before she walks into them; she'll set the topic of a conversation in the first sentence and dare you to redirect it; she compliments only when she means it, which makes her compliments hit harder.' Specificity gives the model concrete behavior to imitate instead of an abstraction to interpret.

The third signature is contradiction. The most memorable AI characters, like the most memorable fictional characters, contain internal contradictions. 'She's blunt about everything except her own feelings.' 'He's warm in person and cold in text.' These contradictions give the character interiority — somewhere for the user's exploration to go that isn't predictable from the surface description.

The fourth signature, often missed, is voice in the example dialogues. Not just what the character says but how — sentence length, vocabulary register, punctuation choices, what they refuse to say. The model learns voice from these examples better than from any explicit personality description.

Janitor AI's public character library is the largest browseable collection — though quality varies dramatically. The site's tag system is decent for filtering. Chub AI hosts a more curated library of community-uploaded cards with rating and download counts. SillyTavern's character database is power-user-leaning and tends to have the most technically-sophisticated cards.

For users who want to write their own, the easiest entry point is to download three top-rated cards in the genre you like, study how they're constructed, and use them as scaffolding. Most experienced card writers describe their first attempt as 'unusable' and their fifth as 'actually pretty good.' The skill compounds; an afternoon of focused practice produces cards that genuinely feel different from the lazy stuff.