How do you keep AI in character? Use a real character sheet, not one adjective

Imagine a TV character who has spent eight episodes speaking in short, guarded sentences. Suddenly she says:

I’m sorry you had that experience. I’ll do my best to assist you.

You immediately feel the break.

AI roleplay can do exactly this. A persona feels convincing for five turns, then slowly becomes polite, generic and agreeable. Rivals forget they were rivals. A blunt character starts over-explaining everything.

The problem is often not solved by typing “stay in character.”

A character is a set of behaviors, not a label

“Cold,” “sweet” and “tsundere” are broad categories.

A stronger instruction explains how the trait appears in behavior.

Instead of:

Personality: cold.

try:

She rarely discusses her feelings directly. When concerned, she solves practical problems instead of saying “I’m worried about you.” She dislikes sentimental language and does not praise people repeatedly. If pressed about emotions, she changes the subject first.

Now the model has patterns it can apply in new situations.

A practical character sheet has seven useful sections

1. Identity

Who is this person? Include age, work and setting only when those facts actually affect the conversation.

2. Core traits

Choose three or four. Explain what each looks like instead of listing ten adjectives.

3. Values

What matters most—freedom, family, efficiency, loyalty, pride, winning?

Values help the character make consistent choices.

4. Voice

Short or long sentences? Emoji or none? Formal or casual? Does the character swear, tease, ask questions, avoid certain words?

5. Relationship with you

“Friend” is vague. Newly acquainted, childhood friend, reluctant ally and almost-dating all produce very different dialogue.

6. Boundaries

What does the character not do?

Examples: never suddenly confesses love, never becomes excessively flattering, never decides on the user’s behalf, never uses customer-service apologies.

7. Current scene

Where are you? What happened immediately before? What does each person want from the scene?

A role feels more coherent when it exists somewhere instead of floating in an endless blank chat room.

Dialogue examples are powerful

If voice matters, give two or three tiny examples.

Me: Were you worried about me?
Character: Don’t flatter yourself. I just don’t want to clean up the mess you’ll make if I leave you alone.

That example turns the abstract instruction “hides concern behind sarcasm” into an actual language pattern.

You do not need twenty examples. A few well-chosen ones can establish the flavor.

Long roleplay needs a recap

Long conversations are limited by the context window. Early scenes may eventually be summarized, compressed or dropped from the material the model can use directly.

For a long-running story, make a short recap every so often:

• Relationship: distrust has become reluctant cooperation
• Important event: survived one crisis together
• Secret: character still does not know the user’s true identity
• Fixed trait: character does not show vulnerability directly
• Next scene: both are waiting at a train station after midnight

This is the roleplay version of “Previously on…”

You are not replaying the entire series. You are putting the current state back on the desk.

Why do dedicated character apps often have memory fields?

Some C-initial character platforms and R-initial companion services expose character memory, relationship state or key memories as product features.

Behind the chat interface, the product may keep feeding the model selected information about the persona and prior events.

That can make a dedicated roleplay product feel more consistent than a general assistant.

But memory is not a perfect recording. Systems can summarize badly, omit a detail or accidentally treat a joke as canon. Important story state is worth checking yourself.

A template you can reuse

Identity:
Core traits:
Values:
Speaking style:
Relationship with me:
Things this character never does:
Current scene:
Two dialogue examples:
Current recap:

Then add:

Keep the persona and current scene as the primary constraints. If a new message conflicts with established character facts, do not silently rewrite the persona.

That is much more useful than repeatedly saying “you broke character.”

Do not make the character perfect

A character who always understands, always agrees and always says the ideal thing becomes flat very quickly.

Habits, blind spots, preferences and things the character refuses to say make the persona more stable—and more interesting.

Consistency does not mean the AI gives the same reaction every time.

It means different reactions still feel like they came from the same person.