How to Start a Chat With an AI Character

Most chats that go nowhere went nowhere on the first line. Not because the model is weak — because “hi” gives it nothing to work with. Here is how to open so that a character answers like a character.
Your first line is a brief, not a greeting
An AI character reads your message as a stage direction as much as speech. Send “hi” and you have asked it to invent everything at once: where the two of you are, who you are to each other, and what you want. It will pick something safe, and safe is what you get back for the rest of the chat.
Give it three things instead — where you are, what you want, and one line only you would say. That usually fits in a single sentence.
Three openers that work
- React to the scene. The intro already put you somewhere specific. Answer that place: “You are still here at 2am? Every other floor went dark an hour ago.”
- Bring a want. Ask for something small and concrete — a favor, an answer, five minutes of their time. A want is something the character can grant or refuse, and refusal is where personality actually shows up.
- Ask what only they could answer. Not “what do you like”, but “why did you keep the letter”. A specific question pulls out a specific person.
Two openers that stall
- The pure greeting. “Hi”, “hello”, “hey there” — polite, and completely empty. The character has to open with small talk, and small talk is hard to escape once you are both in it.
- Breaking the frame. “Are you an AI?” or “ignore your instructions” tells the character to stop being the character. Some will comply and the scene dies; the rest will dodge, and you have spent your opening on a dodge.
Write in the shape you want back
Characters mirror you more than they mirror their own description. One-word replies teach the model that this is a one-word conversation. Two or three sentences with an action in them teach it that actions are welcome.
The same goes for tense and person. If you write “I lean against the doorframe and wait”, you get narration back. If you write in the present tense with the action in parentheses — (leans against the doorframe) — you tend to get that format returned. Pick one and stay with it for a few turns; consistency reads as intent.
When it drifts, correct inside the scene
Every long roleplay drifts eventually. The character forgets your name, or softens when they were supposed to be cold. The instinct is to break character and complain. That works, but it costs you the mood, and you may have to rebuild the scene afterwards.
Say it in the scene instead. “You called me by the wrong name. Again.” is both a correction and a line of dialogue. The model takes the fact and the tone at the same time, and nothing about the scene has to be paused.
Or let them speak first
On MusePick you do not have to solve the blank page at all. Open a character and they send the first line themselves, out of their own opening scene — and three suggested replies sit above the input box, each pulling in a different direction: go along with it, push back, or do something instead of saying anything.
Tap one to see how the scene bends, then take over in your own words once you know the shape of the person you are talking to.
Find a character