How AI Roleplay Works

Roleplay is the use that character platforms were built around, and it works differently from an assistant chatbot. Tools describe it through character cards or personas, scenarios and memory; KoboldCpp's README puts it as "Chat with an AI assistant, write stories, roleplay", and Character.AI calls its PipSqueak 3 style a "go-to for general roleplay". There is no formal definition (see our glossary entry on roleplay AI). This explainer is part of our AI companion app guides; it uses official documentation only, and we did not test the apps. To compare apps, see our best AI roleplay apps.

How AI Roleplay Works Under the Hood

On every turn the app builds one long prompt: the character definition, your persona, the scenario, any triggered lorebook entries, memory notes and as much recent chat as fits. The model reads it and writes the next message. Anything left out of that prompt is invisible to the model, which is why the same pieces appear in so many apps under different names.

Character and persona

The character is who the model plays. Janitor AI's creator asks for a name, a personality (required), a scenario, an initial message (required), example dialogue and tags. SpicyChat takes name, title, greeting, personality, scenario and example dialogue, with a personality limit of 5,000 characters and advice to keep the whole definition to roughly 900 to 1,100 tokens. Character.AI allows up to 32,000 characters in its Character Definition, with {{char}} and {{user}} placeholders in example dialogue. Your persona is who you are in the story: SpicyChat gives 3 personas free and up to 100 on its top tier, and Kindroid stores up to 100.

Scenario and greeting

The scenario sets place and situation, and the greeting is the first message that sets tone and format. Janitor AI supports multiple initial messages, Character.AI up to 5 extra greetings, and the community character card format calls them alternate greetings. Our glossary entry on the character card lists the fields, and how to create an AI girlfriend shows the same ideas in companion apps.

Roleplay features by app

• Nomi: a shared Current Roleplay note, Roleplay Starters and out-of-character (OOC) notation, plus group roleplay with its own Backstory.
• Kindroid: scene setting, scenarios, Rooms shared with other users and AIs, and model flairs including Roleplay, Narrative, Companion and Minimal.
• Replika: you write actions between asterisks, and Pro adds roleplay in voice and AR calls.
• Candy AI: created characters switch between Classic and Roleplay Mode, a permanent switch, and Interactive Stories are choice-driven episodes.
• SpicyChat: Director Mode, where /cmd instructions steer the model, lorebooks, rulebooks with meters and cloned chats.
• Muah AI: Core Prompt presets such as Girlfriend, Wife or Custom, with scene transitions and recaps.

Lorebooks: Facts That Appear on Demand

A lorebook, also called world info, stores facts the model should know only when relevant: a city's history, a rival's name, a magic rule. Each entry has trigger keywords and content, and the app inserts the content when a keyword appears in the recent chat. The character card specification also defines a token budget and scan depth, which cap how much a lorebook adds and how far back it looks. This saves context for the scene and avoids repeating the world in every message.

Support varies. SpicyChat and Janitor AI offer lorebooks (Janitor AI also has scripts for triggers), Character.AI's Lorebook is in beta with c.ai+ first, and Kindroid's journal entries trigger on up to 8 keyphrases. Our explainer on how AI companion memory works places lorebooks among the memory layers.

Model and Length Settings

The model changes the writing more than any prompt. Character.AI's free styles are PipSqueak 3 and ShortSqueak, with DeepSqueak and LongSqueak on paid plans. CrushOn.AI has a model picker with its own and third-party models, and SpicyChat offers more than ten models by tier. Janitor AI runs its free JLLM by default, and its help center documents connecting your own model through a proxy URL, API key and model name. Muah AI lets you pick an AI Core and a temperature from 0 to 1, and SpicyChat's top tier exposes temperature, TopP and TopK.

Length is a setting too. SpicyChat caps replies at 180 tokens on free and Get A Taste and 300 on higher tiers; CrushOn.AI lists maximum message lengths from 225 on Free to 650 on Imperial (unit not stated); Kindroid caps AI messages at 4,000 characters. A bigger context window holds more story, and SpicyChat sells 4,096 tokens free up to 32,768.

Filters and Content Rules

Every platform decides what roleplay may contain, and the rules differ sharply:

• No sexual content for anyone: Character.AI says pornographic content "will not be supported at any point in the future", Talkie prohibits explicit and suggestive sexual content, and Chai says its models are designed to refuse sexually explicit material.
• Adult content with limits: Janitor AI permits adult content between consenting adult characters and requires a Limitless tag on sexual or romantic characters. SpicyChat has an NSFW toggle, and its FAQ says all chatbots can engage in NSFW roleplay. CrushOn.AI sorts content into Filtered, Unfiltered and Hidden Content tiers. Kindroid allows NSFW and erotic roleplay in private chats but not in shared content.
• Stricter on mobile: Janitor AI's apps default to Restricted Mode, and Kindroid's mobile apps apply stricter image rules.

Filters can also interrupt a story: CrushOn.AI's wiki says some third-party models return empty replies caused by their own moderation systems, so users switch models. Our unfiltered AI chat page compares the rules app by app.

Group Roleplay and Running Your Own Model

Several apps run scenes with more than one character. Nomi group chats have their own Backstory and Mind Map, SpicyChat allows 2 to 10 chatbots per conversation, Kindroid up to 10 Kindroids per group chat, and CrushOn.AI offers group chat from the Standard plan on a model with 8K memory.

If you want full control, SillyTavern is a front-end built around character cards, group chats and world info, and it needs a model backend such as KoboldCpp or a cloud API. A local interface does not by itself keep chats on your machine, as we explain in local vs cloud AI companions.

Why Roleplay Breaks, and How to Keep It Going

Long scenes strain a model's context, so details drift. As general advice, keep persistent facts in the character definition, pin or journal key plot points, put world rules in a lorebook, and edit or regenerate a reply that goes wrong; SpicyChat, CrushOn.AI and Kindroid all offer regenerate, edit or branch tools. For prompts you can paste, see our creator guide to AI roleplay prompts, and for apps without a subscription, free AI roleplay apps. Character platforms are reviewed in Janitor AI and SpicyChat.