Writing a first message that gets a usable reply from an AI character

A character card hands you a scene. What you type back decides whether the model continues that scene or quietly replaces it with a generic one, and most flat first exchanges come from that swap rather than from a weak character. What follows is the mechanism as we can read it in Charmi’s own card data and chat pipeline, plus three openers you can adapt and test on any card yourself.

Read the card before you type

Every card carries an opening message, sometimes an opening narration, and often a set of alternate greetings written by the card’s creator. Those greetings are not different personalities. They are different situations, and picking one is picking the premise you have agreed to play inside. Creator notes, where a creator wrote them, usually state that premise directly.

Your first message has one job at this point: continue the situation you chose, and add one concrete thing the character can react to. Cards that open mid-event hand you an obvious hook. Cards that open on a neutral present-day scene hand you almost nothing, so you supply the specificity instead. The AI Roleplay Characters hub collects more of the first kind; AI Companion Characters leans towards the second.

What we observed

On the allowlisted card Audrey, a cafe barista, the creator’s notes list three alternate greetings: 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. Same character, three separate situations. On Kyria, a Queen’s Guard, the main opening is a direct plea for help because her father is ill, while an alternate greeting starts mid-mission in sewer tunnels under a city. Selecting a greeting sets the scene, not the personality.

Under the opening, the product shows two short reply choices, stored on the card as option_a and option_b. These are platform-generated rather than written by the creator, and on SFW cards they are frequently generic flirtatious lines that do not fit the scene at all: Audrey’s card offers to hear what you want tonight, Kyria’s offers undivided attention after the battle. Tapping one is fast, and on a story-first card it pulls the tone away from whatever the creator built. Typing your own opener is the cheap fix, and it is the only part of the first turn you control.

The chat page asks for a display name the first time you talk to a character. That name is substituted into the character’s prompt as {{user}}, so it is what the character calls you from the first reply onward. A placeholder you would not want said out loud in a scene will be said out loud in the scene.

Memory extraction runs after a completed turn rather than during streaming, keyed by turn ID so retries do not create duplicate records. Anything you state plainly in your opening line is therefore already a candidate the moment that first exchange finishes, and the chat surfaces one quiet remembered event with a Review action when a record is proposed. The opening message is the earliest place you can seed memory, which makes it worth spending a clause on something true and durable.

Three openers you can adapt

For a cafe scene where you are a first-time customer:

I’ve walked past this place every morning this week and never come in. What do you make for someone who has no idea what they want?

That accepts the location, hands the character a concrete question she can answer in her own voice, and says one small thing about you.

For a high-stakes scene already in motion:

Keep your voice down, the water carries further than you think. Which way were you heading before I caught up?

Note what is missing: no summary of the premise recited back at the character, no stage directions, no meta commentary about the roleplay.

For a card that opens on nothing in particular:

I’m on a late shift with about twenty minutes free. Tell me about your day first and I’ll tell you about mine, I’m Sam by the way.

That one plants a name and a time constraint. Both are durable facts rather than passing mood, and the constraint gives the character a reason to keep replies short instead of writing three paragraphs of scenery.

What to leave out

Skip the instructions. Lines such as stay in character, be descriptive, do not break the fourth wall spend your opening turn on rules the card already encodes, and they push the model into talking about the roleplay instead of doing it. Skip the backstory dump too. Four sentences about who you are gives the model too many threads and it will usually pick the least interesting one.

Avoid yes-or-no questions. They invite a one-word answer followed by filler. Ask something the character has to answer from a point of view, or state something they can contradict.

Judge the second reply, not the first

One good reply proves little. Send your opener, read the answer, then send a second message that depends on a detail from the first: reference the thing they made you, or the direction they said they were heading. If the second reply forgets it, the problem is continuity rather than your opening, and the tests in how to test an AI character’s consistency will isolate it faster than rewriting the opener again.

When you want more scene ideas to adapt, AI roleplay opening scenarios that sustain a story works through premises that survive past the first exchange, and Character AI alternatives for long-form memory covers what happens to those early facts across a longer conversation. If you would rather start from a card whose premise is already sharp, browse the creators behind them on Creators.