How to Roleplay With AI (2026): What to Type, How Much, and When to Stop

Roleplaying well with an AI comes down to one habit: write what your character tries, and stop there. Everything else on this page is a version of that sentence.

If you searched for how to roleplay with AI properly, or better, you are probably not asking how to start. You have started. Something keeps going wrong — the AI takes over your character, the story wanders, every reply is a wall of text, nothing you do seems to matter — and you suspect it is you. Some of the time it is the platform, and we have a guide to testing that. But a lot of the time it is the message you typed, and that part is completely in your hands.

This guide is about that message: the single turn, on any platform or model. It is not about the setup prompt — how to turn a chat into a game covers that — and it is not about any one app. It works the same in AI Dungeon, in a ChatGPT chat, in Janitor AI or in a dedicated game.

Three Sentences That Break an AI RPG

Almost every bad turn is one of three sentences. Learn to spot them in your own writing and most of the frustration goes away.

1. The sentence that does the game master’s job

I pick the lock, slip into the study and find the ledger hidden under the floorboards.

Three things happen in that sentence, and you decided all three. Did the lock open? Was anyone in the study? Was the ledger there at all? The AI now has nothing to decide, so it has nothing to do except agree and keep writing. Do this for a few turns and the game becomes a story you are writing with a very polite assistant.

The fix is to end on the attempt:

I kneel at the study door and start working the lock, listening for footsteps in the hall.

Now the world has a job. The lock can resist. Someone can come down the hall. The ledger can be somewhere else. That uncertainty is the game, and you gave it back.

This is the most common mistake in AI roleplay, and it is also how you accidentally train the AI to play your character for you. If you narrate outcomes, the model learns that outcomes are narrated, and sooner or later it narrates yours. We explain that failure, and why models fall into it, in railroading and puppeting.

2. The sentence that asks permission

Can I try to climb the wall?

It looks polite, and it quietly turns the AI from a game master into a helper. Helpers say yes. Worse, a question like this invites the AI to answer rather than to resolve — “Sure! You climb the wall and…” — and you are back at sentence one, except now the AI wrote the outcome.

Just try it:

I look for handholds in the old stonework and start to climb.

There is a good kind of question, and it is worth separating from this one. Asking the world what is there is exactly what a player should do: “Is there a way onto the roof from this side?” or “What does the guard look like from here?” Those questions ask for information the game master holds. Asking whether you are allowed to act asks for nothing the fiction needs.

3. The sentence with no verb

ok

…

continue

Each of these hands the pen over completely. The AI has to write something, and with nothing from you it writes its own plot — often including what your character does next. If you have ever wondered why the AI suddenly moved your character across town, check what you typed the turn before.

If you want time to pass, say so, and say what your character is doing while it passes:

I sit in the back of the tavern with a drink, keep my hood up, and watch the door until the courier arrives.

That still lets the world move. It just moves around a character who is doing something.

What a Good Turn Contains

A good turn has three parts, and only the first is required.

1. What your character does or says. An action, a line of dialogue, or both.
2. How, or why — if it matters. “Quietly.” “Keeping my hand near my knife.” “Because I don’t trust him.” This gives the game something to judge the attempt against.
3. A stopping point. You end the turn at the moment the result is unknown.

Here is the same moment written three ways:

Turn	What it does
”I convince the captain to let us sail tonight.”	You decided the result. Nothing is at stake.
”Can I ask the captain if we can leave tonight?”	You asked permission to do something you could just do.
”I put a purse of silver on the captain’s table. ‘Tonight. Before the tide turns. Name your price for the risk.’”	Clear action, clear words, clear stakes — and the captain gets to answer.

The third turn is longer, but that is not why it works. It works because it gives the other character something to respond to and leaves their response to them.

How Much Should You Write?

Less than you think, most of the time.

The AI tends to answer in kind. A short, sharp turn usually gets a tighter reply. A five-paragraph turn usually earns a five-paragraph answer — which you then have to read, and which has had more room to invent things you did not want. If the replies are too long, look at your own messages first, before you add “keep responses short” to a prompt.

A rough guide that works on most platforms:

• In action — a fight, a chase, a heist: one to three sentences. Pace is the point. Each turn should be one decision.
• In conversation: as long as what your character would actually say, plus a short action if it matters. People rarely speak in paragraphs.
• At a scene change, or when you set something up: longer is fine. This is where detail pays — where you are, what your character wants, what they are carrying into the scene.

One thing to avoid at any length: stacking several actions that depend on each other. “I sneak past the guard, open the gate and signal the others” asks the game to resolve three uncertain things at once, and it will usually just let all three succeed. Do the first one, and see what happens.

Dialogue and Action: Keep Them Apart

The AI needs to know which words your character says and which things your character does. The most common convention is:

• Speech in quotation marks: “Put the sword down.”
• Actions as plain narration, or between asterisks: *I step between them* — or simply, I step between them.

Which one you pick matters less than keeping to it. Mixing conventions mid-scene is one of the easy ways to get the AI confused about whether your character spoke a threat or just thought it.

Some platforms build this choice into the input box. AI Dungeon has separate modes: Do for actions, Say for dialogue, Story for narration. In Do mode, its help centre says, “anything you type is recognized by the AI as an action performed by your character,” and the game rewrites “I” into “You” for you. Story mode is different — it adds your text to the adventure unchanged, so you can “force story events to happen” or write from another point of view. That is a powerful tool and, used every turn, it is sentence one with extra steps. Use it to steer, not to resolve your own attempts.

AI Dungeon’s help centre also has a small tip that applies everywhere: vary how your turns start. If every message begins “I…”, the replies drift into the same rhythm too.

First person or second person? Either works. “I draw my sword” and “You draw your sword” both read clearly to every current model. The only rule is consistency — pick one and keep it, and match whatever your platform converts to.

Never Write the Other Character’s Reply

I tell the guard I’m the duke’s nephew, and he believes me and waves me through.

This is sentence one, but it is worth its own section because it does a second kind of damage. You did not only decide the outcome. You spoke for another character. The guard never got to be suspicious, or bribable, or loyal to the real duke. You turned a person into a door.

The fastest way to lose the characters that make an AI RPG worth playing is to write their lines for them. A character only feels real when they can refuse you — and they cannot refuse a sentence you already finished. Say your part and stop:

“I’m the duke’s nephew. He’s expecting me.” I hold the guard’s eye and don’t slow down.

What the guard does next is the reason you are playing. For why AI characters are good one-on-one and blur together in groups, see how AI RPG NPCs work.

Asking the World Versus Telling It

There is a quieter version of the same problem:

I check the back of the tavern for the secret door.

What secret door? If nobody mentioned one, you just created it. In a freeform story that can be fine — some players enjoy co-writing the world as they go. In anything that is meant to be a game, it removes the thing you were supposed to find out.

Ask instead:

I walk around the back of the tavern, looking for another way in.

Now the game decides whether there is one. If you enjoy inventing the world as you go, that is a legitimate way to play — just know you are choosing a story over a game, and the challenge will fade to match. Our difficulty analysis explains why an AI game master that agrees with everything stops being fun.

When to Step Out of Character

“OOC” means out of character — a note to the game master rather than a line in the story. Players usually mark it with brackets:

(OOC: my character is lying here — she doesn’t actually know where the map is.)

Some platforms build this in. Friends & Fables added an OOC mode in March 2025: start a message with “OOC:” or with its game master’s name, Franz, and the patch notes say he is “much more likely to complete your requests accurately” — checking character sheets and memories rather than carrying on the story. Elsewhere, brackets are the convention every model understands.

Use it sparingly, and for jobs the fiction cannot do on its own:

• Flag intent the AI cannot see. Lies and bluffs are the classic case. The narrator only knows what you typed; an unflagged lie can be read as the truth, and then the story writes your character meaning it. We recommend exactly this note in our Voyage combat guide, because it costs a few words and saves the scene.
• Correct a fact. “(OOC: we left the horses at the inn, not the gate.)” Faster and cleaner than arguing about it in the story.
• Set the pace. “(OOC: can we skip ahead to the evening?)” or “(OOC: slow down, I want to play this conversation out.)”
• Fix a rule the AI broke. If it narrated your character’s words or feelings: “(OOC: please don’t write my character’s dialogue — describe the scene and let me respond.)” Correct it the first time it happens. An uncorrected slip becomes a habit.

What OOC is not for: negotiating a result you did not like. “(OOC: can the lock open anyway?)” is sentence two again, in brackets.

When the Reply Goes Wrong: Reroll, Edit or Answer?

Most platforms give you a way to throw a reply away — retry, regenerate, undo, a swipe. It is tempting to use it every time something goes badly. It is worth knowing when to use it and when not to.

Reroll when the reply broke the rules of play. It wrote your character’s dialogue. It contradicted something established a few turns ago. It ignored what you actually did. It refused for no reason — our guide to why AI refuses to roleplay covers that case. These are bugs, not outcomes, and throwing them away is exactly what the button is for.

Answer in the story when the reply is just bad news. The guard did not believe you. The climb failed. The ally said no. That is the game working. Rerolling until you get the result you wanted is save-scumming, and in an AI RPG it has a hidden cost: once you know every failure can be erased, none of them mean anything, and a game where nothing can go wrong gets boring fast.

Some platforms build that judgement into the button. NOPOTIONS, whose engine rolls the dice, lets you undo only your most recent turn, does not refund the credit you spent on it, and will not undo a turn where your character died.

Edit your own message when the problem started with you. If you wrote sentence one and the AI ran with it, fix your turn, not its reply.

When to End a Turn, and When to End a Session

End a turn at the moment of uncertainty. Right before the result — the swing, the question, the step onto the ice. If you notice yourself writing what happened next, delete it.

End a session at a scene break, not in the middle of one. Stopping mid-fight means coming back to a scene whose tension has gone. Stopping when the characters have reached somewhere, or a day has ended, gives the next session a clean start. If you play in a plain chat or a long-running session, a short recap at the top of your next message helps the model find its place — the RPG chat guide covers that habit, and how to summarise a long campaign covers the longer version.

Does Any of This Depend on the Model?

Less than people think. Everything above works on every platform and every major model, free or paid, because it is about what you give the AI to work with, not about how clever it is.

The model and platform do change how much help you get. Some dedicated games hold the rules outside the model, roll dice you can see, and refuse to play your character no matter what you type. Others do exactly what your turn tells them to, which means everything on this page matters twice as much. If you want the specific setup for one tool, we have step-by-step guides for Claude, ChatGPT, Gemini, Grok and Janitor AI. If you have not picked a platform yet, the best AI roleplay platforms is the place to start, and the best free AI roleplay if you do not want to pay.

A One-Minute Checklist

Before you press enter:

• Did I write what my character tries, not what happens?
• Did I ask permission for something I could just do?
• Did I write another character’s words or reaction?
• Did I invent a fact the world never gave me?
• Is this one decision, or three stacked together?
• Can the AI tell what I said from what I did?

Frequently Asked Questions

How do you roleplay with AI properly? Write what your character tries, and stop before the result. Say what they do or say, add how or why if it matters, and end the turn at the moment the outcome is unknown. Don’t write other characters’ replies, don’t invent facts the world never gave you, and keep speech and actions clearly separate. The AI can only make the game interesting if you leave it something to decide.

How long should my messages be in AI roleplay? Usually shorter than you think. The AI tends to answer in kind, so a long turn earns a long reply. In action scenes, one to three sentences per turn works well — one decision at a time. In conversation, write what your character would actually say. Save longer turns for scene changes and setup, where detail helps.

Should I write in first person or second person? Either works on every current model. What matters is keeping to one. Some platforms convert for you — AI Dungeon’s Do and Say modes turn “I” into “You” automatically — so it is easiest to match whatever your platform does.

How do I stop the AI from writing my character’s actions? Start by not narrating outcomes yourself, because the AI copies the pattern. If it still happens, correct it the first time with a short out-of-character note, such as asking it not to write your character’s dialogue and to let you respond. Reroll any reply that puppets your character rather than playing on from it.

What does OOC mean in AI roleplay? OOC means out of character: a note to the game master rather than part of the story, usually written in brackets. Use it to flag a lie the AI cannot see, correct a fact, set the pace, or ask the AI to stop breaking a rule. Don’t use it to negotiate a result you didn’t like.

Is it cheating to reroll AI responses? Not when the reply broke the rules — wrote your character’s lines, contradicted established facts, or ignored your action. Those are errors, and rerolling them is what the button is for. Rerolling every bad outcome until you get a good one is different: it removes the stakes, and a game where nothing can go wrong stops being fun.

Why does the AI keep taking over the story? Usually because the turn before gave it nothing to do. Messages like “ok” or “continue” hand the story to the AI, and messages that decide outcomes teach it that outcomes are decided in the text. Give it an action with an uncertain result every turn, and it has a reason to wait for you.

Product names and trademarks are the property of their respective owners and are used for identification only. Arcanum is not affiliated with any platform named here. AI Dungeon’s input modes quoted from its help centre, 13 September 2026.