Reading an anime AI character card before you start

Most people evaluate an anime character card by its cover art and its first two sentences, then wonder several turns later why the conversation went somewhere flat. The cover is the least informative part of the record. A card is a structured document with named fields, and each field does a specific job when the prompt is assembled. Reading four of those fields before you type anything is the difference between choosing a character and being assigned one.

The field names below come from Charmi’s live character API, but the shape is common across the ecosystem. The widely used Character Card V2 specification defines much the same set — description, personality, tags, creator notes, a first message, and an array of alternate greetings — and tools such as SillyTavern render them the same way. Learn the shape once and it transfers.

What we observed

Reading Charmi’s character API on 2026-09-06, a card carries name, tags, bio, description, opening_message, opening_narration, an alternate_greetings array, creator_notes, option_a, option_b, nsfw_level, and counters for stars and chats. Four of those fields are written by the person who made the card and carry most of the usable signal: tags, creator_notes, opening_message, and alternate_greetings.

Two allowlisted cards show how that plays out. On Audrey, a cafe barista written by the creator thecooler, the notes list three greetings, and the interesting property is that all three are the same person in different situations: a first-time customer at the cafe, a coworker invited to look at records after closing, and an evening at her apartment while she nervously prepares to play a song. Three scenes, one personality. Picking a greeting on that card selects a starting situation and a relationship distance, not a different character.

On Kyria, a Queen’s Guard written by the creator skyrock, the main opening_message is a direct plea for help because her father is ill, while an alternate greeting drops you mid-mission in the sewer tunnels under a city. The second one begins in motion, with the premise already underway, so it rewards a reader who wants to act rather than be introduced. Kyria’s creator notes state the premise plainly — a rival who nearly got her dismissed, and a father with an undiagnosed disease — which is information the opening scene alone does not give you.

The option_a and option_b fields behave differently. They are two short first-reply choices the product renders under the opening, they are platform-generated rather than written by the creator, and on SFW cards they frequently read as generic flirtation that does not match the scene at all. Both cards above show that mismatch. They are not part of what the writer built, and treating them as authorial intent is a reliable way to misread a card.

Tags carry two kinds of information

A tag list mixes genre with rating, and the two are worth separating as you read. Genre tags — school, fantasy, mecha, slice of life, tsundere — describe the register the writer was aiming at. Rating tags sit alongside a numeric nsfw_level, and they tell you how the catalogue filed the card rather than what any given session will produce.

Anime cards in particular lean on archetype tags as shorthand. That shorthand is useful for narrowing a listing on the Anime AI Characters hub, and much less useful as a prediction of conversational depth. An archetype tag tells you the silhouette. The description and notes tell you whether anything was written underneath it.

Creator notes are the closest thing to a manual

Notes are optional, and their presence is itself a signal. A writer who took the trouble to explain the premise, list the greetings, or say which scenes work usually also wrote a description with enough specificity to hold up over a long session. Kyria’s notes are a good example of the form: two concrete plot facts that give you something to ask about immediately.

Where notes are absent, you are reading blind and should weight the opening scene more heavily. Where notes contradict the tags — a note describing a slow-burn mystery on a card tagged for something else entirely — trust the notes, since they are the part a person wrote last and deliberately.

Greetings are scene selection

Treat alternate_greetings as a menu of entry points into one continuity. The choice sets three things at once: where you are, how well the character already knows you, and what the scene expects you to do first. Audrey’s cafe greeting starts you as a stranger; her apartment greeting starts you as someone already trusted. Neither changes who she is.

Beginning mid-action, as Kyria’s tunnel greeting does, suits a reader who prefers to establish their own footing. A cold introduction suits someone who wants the character to do the setup. Neither is better, and the mismatch between what you picked and what you wanted is a common reason a session feels wrong for reasons nobody can name afterwards. There is more on choosing entry points in AI roleplay opening scenarios.

Skip the canned first replies

Since the two suggested replies are generated by the platform rather than written into the card, they carry none of the scene’s information and can pull the register somewhere the opening never pointed. On a story-first card the practical rule is simple: type your own first message. One or two sentences that answer the opening on its own terms will keep the setting the writer established.

Writing that first line well is a skill worth building — how to write a believable AI character personality covers the same instincts from the author’s side, and they apply just as well when you are the one replying.

A two-minute read before the first message

Open the card and go in order: tags for genre and rating, notes for premise, greetings for entry point, description for whether anything specific was written. Then decide whether the premise is something you want to spend an hour inside, and pick the greeting that matches the role you want to play.

If a card passes that read and still goes vague after ten turns, the problem is usually the description rather than your prompting, and testing AI character consistency describes how to check that deliberately. Browsing AI Roleplay with these four fields in mind narrows a very large catalogue quickly.