How to Roleplay with AI (2026 Guide)

AI roleplay has one open secret: the model mirrors you. Give it a flat setup and one-line replies and it produces a flat scene in one-line replies. Give it a specific place, a want, and something in the way, and it produces fiction. This guide is the craft half — what you type, when, and why — with the tool half kept honest: some of this works anywhere, some of it needs an app built for stories.

Start with a scene, not a greeting

The weakest opening is "hi." The model has nothing to work with, so it fills the gap with pleasantries, and pleasantries are where scenes go to die. Open inside a moment instead: where you are, what your character wants in the next five minutes, and the obstacle in the room. One sentence of each is enough. You're not writing a novel's first page — you're handing the writer a situation that can only go somewhere.

If the app has worlds or starter scenes, use one. A world is a context prompt with a cast — it spots you the setting and the stakes so your first line can already be a move.

Play your character; direct everything else

There are two registers, and mixing them up is the most common beginner mistake. In character, you speak and act as yourself-in-the-story. Out of character, you direct: what the narrator should do, where the scene should go, what to slow down. Chat apps only really give you the first register. Story apps give you both — in Underfiction you can speak, act, or step outside the scene and tell the narrator what happens next.

Direction is where the craft lives. "She opens the letter." "Slow down. Let the silence do the work." "Cut to morning." Short, specific prompts win — a director doesn't explain the whole film, they call the next shot.

Push the story

The best scenes come from risks. Let your character make the wrong choice. Say the thing that can't be unsaid. A scene where nothing can go wrong is a screensaver. If the app you're on punishes risk — a filter that trips on tension, a model that lectures instead of writing — that's not a craft problem, that's a tool problem, and no amount of prompt skill fixes it.

Manage the memory, don't fight it

Every model has a context window; what falls out of it is gone. On chat platforms the plot quietly evaporates around turn thirty. The fix is structural: play in scenes. End a scene when its question is answered, start the next one clean. Apps built for long stories summarize past scenes into chapter memory automatically — Underfiction does this in the background — so chapter twelve still knows what chapter two set up. If your app doesn't, keep a running note of established facts and paste it when things drift.

Set standing rules once

Anything you find yourself repeating — point of view, tense, how much the narrator writes for you, a rule like "never speak for my character" — belongs in a standing instruction, not in every message. Underfiction ships these as narration nudges (person, tense, inner thoughts, who writes your words) plus custom directions in your own words, saved with the story and applied every turn. Elsewhere, put them in the character definition or system field. Rules you state once hold; rules you nag about decay.

Common failure modes, quickly

• The bot speaks for you — set an agency rule ("my character's words are mine"); apps with an agency control enforce it.
• Every reply is a wall of text or three words — length is steerable; ask for it or set a length nudge.
• The scene races to the ending — pacing direction: name what should not resolve yet.
• Everyone sounds the same — the character needs example dialogue, not more adjectives.
• It forgot the plot — scene structure and summaries, not longer messages.