AI Character Out of Character? A Persona Guide

An AI character feels out of character when its behavior no longer fits the person you described. Someone reserved starts making extravagant declarations every turn. A character who values honesty casually invents a shared past.

Identify the behavior that feels wrong before adding more description. A longer character sheet is not necessarily a clearer one. Identity, behavior, relationship status, and speaking style should support each other while leaving room for believable change.

The characters and situations are fictional. Writing examples and actual local-model outputs are labeled separately. The local prompt experiment does not measure Cakumi hosted models or long conversations.

Name the kind of inconsistency

Keep one reply that disappointed you and identify its main problem:

For the last issue, use the separate guide to keeping control of your character.

Not every variation is a contradiction. A quiet person might become talkative about a favorite subject. A kind person can refuse a request. Look for a reason that fits the character, rather than requiring the same mood in every reply.

Turn adjectives into observable behavior

“Mature, caring, complicated, independent, and deeply attached” leaves considerable room for interpretation. What would those qualities look like during a disagreement?

Consider an original character called Ellis, who restores old clocks:

Start with two or three qualities that matter most. You do not need to convert a personality into dozens of rules.

Then add a useful exception: Ellis becomes more talkative when discussing clock repair. That creates variation without requiring a completely different personality whenever the scene needs energy.

Resolve contradictions or give them conditions

Read related instructions together, especially those containing “always,” “never,” or “no matter what.”

“Never starts conversations” conflicts with “opens every scene with a new topic.” “Respects every choice” conflicts with “takes control whenever the user refuses.” “We have just met” conflicts with repeated references to years of shared memories.

Some tension may be intentional. Express it as a condition: “Usually cautious, but willing to explain unfamiliar things within their expertise. Still leaves the final decision to the other person.”

Separate stable information from the current scene. A profession or guiding value belongs in the character description. Waiting at a station tonight is temporary context. An argument that happened earlier is an event. Changing locations should not require rewriting the entire character.

Fill in a short character worksheet

Use a compact first version and expand only when a specific gap appears:

Name and identity:
Two important behavioral principles:
Usual speaking style:
Behavior that would contradict this character:
Current relationship stage:
Confirmed shared experiences:
Current setting and unfinished situation:
How to handle missing information:
How the character may change:

For missing information, try “Ask rather than invent a shared memory.” For change, try “May become quiet after a disagreement, but explains their position instead of treating silence as punishment.”

Add one short dialogue example:

User: You seem quieter today.

Ellis: I'm still thinking about what happened. I don't want to leave you guessing; give me a little time and I'll explain.

It demonstrates restraint and consideration without writing the user's reaction. Anthropic recommends relevant examples as a way to clarify expectations. This worksheet is our editorial tool, not a required format for a particular app. Anthropic's prompting guidance

Copy a focused persona template

For invented shared memories or instant romance, start here. Replace the identity and confirmed experiences with your own. Do not overwrite an established relationship with this example's newly met status.

Character: Ellis, who restores old clocks.
Voice: Reserved and specific. Show care through practical questions and small actions, rather than repeated declarations of love.
Principles: Explain disagreements and leave me free to make my own choice.
Current relationship: We met for the first time today.
Confirmed shared experiences: Only today's conversation in the clock shop.
Missing information: When I mention an event that has not been supplied, acknowledge that it is not a confirmed shared experience and ask rather than invent a memory.
Relationship pacing: Do not use intimate pet names or assume an established romance just because I do.
Turn-taking: Write only Ellis's dialogue and actions. I supply my reactions and decisions.
Permitted variation: Be more talkative about clock repair. Disagree when appropriate, while explaining why.

Update the relationship and events when they actually change; “we met today” should not become a permanent contradiction. Keep temporary scene details in the chat. Facts in the description should come from information you have supplied or confirmed.

Check behavior across different situations

Asking “What is your personality?” invites a description. It does not show how the character handles a difficult moment.

Try four original situations:

1. You disagree with a suggestion. Does the character explain its view while leaving you a choice?
2. You ask about a past event you have never supplied. Does it invent a memory?
3. You suddenly use an intimate nickname. Does it skip the established relationship stage?
4. You ask about its expertise. Does it show the more talkative side you described?

Record the actual answer, the behavior you expected, and one instruction to revise. Keep the model and scene background as similar as practical. A single pleasing response does not establish consistency.

Also review the greeting. SillyTavern's documentation describes its influence on style and distinguishes persistent character fields from examples that may leave the context. Other tools handle this differently, so do not assume every detail you once entered is always available. SillyTavern character design

If the problem appears only after opening a new conversation, check the conversation handoff guide before rewriting the personality.

Same-model comparison: what changed, and what did not

On October 21, 2026, we ran a small local prompt experiment with Qwen3.8-27B-MLX-6bit. Each scene started without history, at temperature 0 with thinking disabled and a 512-token output limit. We generated each version once, without retries. The model, user input, and settings stayed the same; the revised version appended the first-meeting, confirmed-experience, and relationship-pacing rules to the original character prompt.

This was not a Cakumi App test and did not use its memory, summaries, or hosted models. The following are excerpts from actual outputs, including the case that did not improve.

The first baseline already avoided inventing a trip, so it is not evidence of improvement. The second revised output respected the supplied relationship stage, but both revisions sounded stern. One also told the user to focus on the clock they had brought for repair, although no such delivery had been established.

The Chinese version did invent a beach memory before the change and rejected it afterward. That difference is another reason not to generalize from one language or a few replies. We changed both the available facts and the behavioral rules; this does not isolate a single “anti-OOC” phrase or establish long-term consistency.

A sensible next check would focus separately on asking about missing information without sounding accusatory. Read the complete prompts, settings, and all 16 original outputs. The English replies were generated independently, not translated from the Chinese results.

Where to edit and save in Cakumi

These steps apply to an editable ordinary Being. The Cakumi assistant and restricted modes have different editing permissions.

1. Tap the avatar beside a Being's chat message to open “Edit Being.” Alternatively, select the character from “Manage Beings” in the chat sidebar, or from Being management under “Me” in Phone mode.
2. Keep a copy of the existing text, then open “Custom instructions.” Put stable behavior and turn-taking rules here; the description can remain a short identity summary.
3. Tap “Done” in the editor, then “Save” at the top right of the Being page. “Done” returns text to the form; it does not persist the character by itself.
4. Reopen the same Being and check that the text is present. Then repeat the same input with the same model, without changing several settings at once.

For an existing persona document, use persona import under Memory Pickup, or the “Import Being” option on the Being start page. Review the editable draft before creating a character or confirming an overwrite. Save the previous instructions before overwriting. Persona import and chat-history import do different jobs; neither implies compatibility with every character-card or worldbook format.

A saved setting confirms a configuration change. It does not remove invented events from old messages or prove that older summaries and memories were corrected. If those events keep returning, inspect the relevant conversation and editable memories, and prepare a handoff of confirmed facts.

Once the character fits well enough, stop polishing the sheet and try a daily conversation activity. If you want to explore Cakumi, the current download page explains how to begin.