In AI roleplay, “OOC” means out of character: a message where you speak as yourself, not as your character, to correct the story, set a boundary or give a technical instruction. When the AI itself goes out of character, the fix is almost always the same: a short, clearly marked OOC note, a reminder of who the character is, and, if the chat has grown long, a fresh summary of the scene. This guide shows exactly how to do all three.
TL;DR
- OOC is you talking to the AI about the story. Mark it clearly (parentheses or brackets with “OOC:”), keep it to one or two sentences, and then return to the scene.
- Characters drift because of long chats, vague profiles, and a model that treats your latest message as the strongest signal. Most of it is fixable.
- Tell the AI what to do, not only what to stop doing. “Stay guarded and sarcastic” works better than “don’t be nice.”
- If one OOC note does not hold, restate the character in a short recap, edit or regenerate the bad reply, and tighten the character profile.
- Use OOC for boundaries too. A good adults-only platform should respect them.
Table of Contents
- What “out of character” means in AI roleplay
- When an OOC message is worth sending
- How to format an OOC message so the AI treats it as an instruction
- OOC examples you can copy
- Why AI characters go out of character
- How to get a character back in character
- Build characters that stay in character
- Our perspective: OOC is a safety tool as much as a style tool
- Build a character that holds its voice
- A quick OOC checklist
- FAQ
- Sources
What “out of character” means in AI roleplay
Roleplay has always had two layers. The in-character (IC) layer is the story: dialogue, actions, thoughts, all written as the character. The out-of-character (OOC) layer is the real people, or in this case the real person and the model, talking about the story.
The habit comes from tabletop games and forum roleplay, where players needed a way to ask a rules question or say “let’s skip to morning” without a character saying it. AI chat inherited the convention because it solves the same problem. You are the author and co-star; the AI is the other half of the cast. Sometimes the author needs to speak.
There are really two situations people mean by “AI out of character”:
- You step out of character on purpose. You send an OOC message to steer the plot, fix a mistake, or set a limit.
- The AI steps out of character by accident. A sarcastic rogue suddenly sounds like a customer-service agent, a medieval knight mentions smartphones, or the character starts lecturing you in a flat, generic tone.
The tools for the first situation are the fix for the second, so we will cover both.
When an OOC message is worth sending
Not every wobble needs one. Swiping or regenerating a reply is quicker when the answer is merely bland. Reach for OOC when something needs to change going forward:
- Corrections. The character forgot your name, the setting, or what happened two scenes ago.
- Direction. You want a slower pace, a time skip, a different tone, or a new scene.
- Boundaries. There is a topic, intensity level or kind of content you do not want in the story.
- Technical instructions. Shorter replies, no speaking for your character, a particular point of view.
- Meta questions. “What do you remember about the village?” is a way to check whether the model still holds the details you rely on.
If you are not sure whether something warrants a note, ask whether the next ten replies would improve if the AI knew it. If yes, send it.
How to format an OOC message so the AI treats it as an instruction
The model has no built-in switch for “this is the author talking.” It reads everything as text, so your job is to make the boundary between story and instruction obvious. Four habits help.
Use a consistent marker
Pick one format and never vary it. The most common are:
(OOC: ...)[OOC: ...]// ...or(( ... )), borrowed from forum roleplay
The exact brackets matter less than consistency. When every instruction looks the same, the model learns that this pattern means “step outside the scene,” and that signal strengthens as the chat goes on.
Put the instruction on its own
Do not blend an OOC note into a paragraph of dialogue. Send it as a separate message, or set it apart on its own line at the start or end of your turn. Mixed messages produce mixed results, such as a character reading your instruction aloud as if it were speech.
Keep it short and concrete
One or two sentences. Name the behavior you want, not a feeling. “(OOC: Keep Mara cold and formal with strangers. Short sentences, no jokes.)” is far more usable than “(OOC: Mara is acting wrong, please fix.)”
Say how to return
End with a cue back into the scene: “Resume from her last line.” Without it, some models keep talking about the story instead of continuing it, and you end up in a long meta conversation.
OOC examples you can copy
Adapt names and details, but keep the structure: marker, specific instruction, way back to the story.
Correcting a character
(OOC: Dana is a former detective and would never say that. Keep her dry, observant and a little suspicious. Continue from her last line.)(OOC: You wrote that my character agreed. I didn’t. Please redo the reply without deciding my actions.)(OOC: Reminder: it is 1923 and we are on a ship. No modern slang or technology.)
Setting boundaries
(OOC: I’d like to keep this scene non-violent. Let tension build through dialogue instead.)(OOC: That went further than I wanted. Pull back to a gentler tone and don’t return to it unless I ask.)(OOC: Pause the story. I’m not comfortable with this direction. Let’s move to a different scene.)
Boundaries are the most important use of OOC, and you never need to justify one. If a platform keeps pushing past a clear limit, that is a product problem, not a failure in your wording.
Steering a scene
(OOC: Skip ahead to the next morning. Summarize the night in two sentences, then start the scene at breakfast.)(OOC: Slow this down. Describe the room and what Leo notices before anyone speaks.)(OOC: Introduce a complication: a phone call interrupts them. Don’t resolve it yet.)
If you want more ready-made starting points, our AI roleplay prompts guide has 30 ideas and a formula you can adapt into scene openers.
Technical instructions
(OOC: Replies of two to four paragraphs, please. Third person, past tense.)(OOC: Don’t write dialogue or actions for my character. Stop at the moment I would respond.)(OOC: Before continuing, list in three bullets what you know about the setting.)
That last one is a quiet diagnostic. If the bullets are wrong, you have found the real problem, and you can correct the record in one message.
Why AI characters go out of character
Understanding the cause tells you which fix to try. Several things tend to be behind it.
The chat outgrew the model’s working memory. Language models read a limited window of recent text. In a very long roleplay, early details, including the character’s original personality, can fall out of view or get less weight than the most recent exchanges. The character then regresses toward a generic, polite default.
The character profile is thin. “Confident, flirty and mysterious” gives the model three adjectives and a lot of room to improvise. A profile with specific speech habits, history, goals and things the character would never do gives it something to hold onto.
Your latest message outweighs everything else. Models weigh recent instructions heavily. If you write casually or in a different tone, the character may mirror you. If you write a long OOC rant, the model may mirror that too.
Safety and style layers sometimes override the persona. Most chat models are trained to prioritize certain instructions over others. OpenAI has described an instruction hierarchy approach that trains models to follow higher-priority instructions and selectively ignore lower-priority ones. In practice, this means a character profile or an OOC note can lose out to higher-level rules and defaults, which is why a character sometimes suddenly sounds like an assistant. Clearer, better-placed instructions help, but they cannot override every layer.
Contradictory signals. If your profile says “shy” but your scenario rewards boldness, or an earlier OOC note conflicts with a later one, the model averages them and you get a muddy character.
Your OOC was read as dialogue. A vague or unmarked note can be treated as part of the story, so the character answers it in voice instead of obeying it.
How to get a character back in character
Work through these in order. Each step costs more effort than the last, so stop when the character recovers.
- Regenerate or edit the bad reply. One off-key message left in the history teaches the model that off-key is normal. Delete or rewrite it rather than correcting it afterward whenever your app allows.
- Send a short OOC reminder. Name the character, two or three defining traits, and the current situation. Then tell it to continue.
- Describe the target, not just the problem. “Stop being so polite” leaves the model guessing. “Speak bluntly, interrupt people, never apologize” gives it a voice to adopt.
- Give a mini recap. For long chats, paste a five-line summary: who is present, where you are, what just happened, what the character wants, and how they speak. It puts the essentials back in front of the model.
- Show an example line. One sample of the voice, such as “She’d say something like: ‘Spare me. Where were you at nine?’”, is often more effective than three sentences of description.
- Update the character profile. If the same drift keeps coming back, the profile is the cause. Add the missing traits and “never” rules there so you do not need to repeat them in every chat.
- Start a new scene with a carried-over summary. When a chat is very long, a fresh start with a tight recap often beats trying to repair the old thread.
A small but powerful habit is to pin or repeat your key rules at the top of the profile or opening message, in plain language, and keep them short. Ten clear lines outperform a wall of text.
When the character ignores your OOC entirely
If the model keeps answering your OOC in voice, try three things. First, change the marker to something more distinctive, such as [[OOC: ...]]. Second, make the note the only thing in the message. Third, phrase it as a positive instruction with a concrete next step. If it still fails, the platform’s defaults may be overriding your note, and the fix lives in the profile, not the chat.
Build characters that stay in character
Prevention beats repair. A character that is written well needs far fewer OOC notes. These elements do most of the work:
- Voice samples. Two or three example lines in the character’s actual speech.
- Concrete traits with behavior. Not “brave,” but “steps between people in arguments and hates being thanked.”
- Goals and fears. They give the character reasons to act, which keeps responses consistent.
- Hard limits. What the character would never say or do.
- Setting rules. Era, tech level, location, relationships.
If you want a deeper walkthrough, read our guide on how to create AI characters that stay in character. For a broader introduction to the craft, see how to roleplay with AI. Writers who like to study how character sheets and prompts steer in-character responses in other kinds of persistent worlds can also look at character profiles and prompts as a different example of the same principle: clear identity in, consistent behavior out.
Our perspective: OOC is a safety tool as much as a style tool
Most advice treats OOC as a way to fix plot problems. We see it differently. On an adults-only platform, where stories can be intimate, intense or emotionally real, the ability to step outside the scene is how a reader stays in control of the experience. It is the roleplay version of a safe word: always available, never needing an explanation.
That shapes how we think about character design. A well-built character should respond gracefully when you step out, acknowledge the note, and return to the story without sulking or pushing back. It should not treat your boundary as part of the plot to be argued with. When we talk about immersive storytelling with customizable characters, privacy safeguards and a curated adults-only environment, this is what we mean in practice: you set the terms, and the story adapts.
We also think the best fix for drift is rarely a longer OOC message. It is a sharper character: a few specific traits, a clear voice and clear limits, set up once and refined as you learn what your story needs.
— Chatr Team
Build a character that holds its voice
If you are tired of repairing the same character every few messages, start at the source. On Chatr you can create your AI girl with the personality, speech style and backstory you want, then chat with her in an adults-only space built around roleplay and storytelling. You can also try creating a companion from one photo if you would rather begin from a look and refine the personality from there. Chatr is available on iOS and Android, and you can start free and decide later whether the premium features are worth it.
Spend ten minutes on the profile, with voice samples and hard limits included, and you will send far fewer correction notes later.
A quick OOC checklist
Before you hit send on your next note, run through this:
- Is it clearly marked as OOC, in the same format as before?
- Is it one or two sentences?
- Does it say what to do, not only what to avoid?
- Does it name the character and the situation?
- Does it tell the AI how to resume the scene?
If all five are yes, you have given the model the best possible chance to follow it.
FAQ
What does OOC mean in AI chat?
OOC stands for “out of character.” It marks a message where you speak as yourself, not as your character, usually to fix a detail, change direction, set a boundary or give a technical instruction. It is usually written in parentheses or brackets, such as (OOC: skip to morning).
Why does my AI character keep ignoring my OOC messages?
Usually the note is unclear, mixed into dialogue, or buried under a very long chat. Send it alone, keep it short, use the same marker every time, and state the behavior you want. If it keeps failing, strengthen the character profile, because defaults and higher-priority instructions can outweigh a single chat message.
How do I make an AI stay in character during a long roleplay?
Start with a detailed profile that includes voice samples and “never” rules. Every so often, send a short recap of the scene and the character’s traits, and delete or regenerate replies that drift. When a thread becomes very long, begin a new chat with a summary.
Can I use OOC to set boundaries?
Yes, and it is one of the best uses. Say plainly what you do not want and what to do instead, for example “keep this non-violent and move to a different scene.” You do not owe the story an explanation, and you can pause or end any scene at any time.
Is there a difference between OOC and a system prompt?
Yes. A system prompt or character profile sets the standing rules for every chat. An OOC message is a one-off instruction inside a single conversation. Use the profile for lasting traits and OOC for adjustments in the moment.