How to keep AI characters in character
Updated
AI characters drift when the conversation pulls them away from their written voice: an out-of-character reply left in the chat, messages that treat them like an assistant, or a thin personality. Fix the first drifted reply instead of continuing past it, and give the character specific traits and "never" rules. On UnsavedAI, Edit and Another version make that a few seconds' work.
Try it now, free, no account neededStart a story with Raven SinclairShe finds everyone tedious. You, marginally less.What drift looks like
The cynical detective starts apologising. The deadpan scholar starts using exclamation marks. The villain asks how they can help. Drift is rarely sudden: one softer reply goes unchallenged, the model reads it as the new normal, and the next reply is softer still.
The earliest sign is usually vocabulary. If Raven Sinclair says "Aww, that's so sweet!", she has already lost the thread. Her version is: "That was almost pleasant. Do not make a habit of it."
Fix the reply, not the next one
Every reply in the chat teaches the model who this character is. So the most effective fix is to remove the bad example:
- Tap Another version and pick a take that sounds right.
- Or edit the reply: rewrite the one line that broke voice, keep the rest.
- If several replies drifted, Undo back to the last good one rather than patching each.
Arguing with the character in the story ("You'd never say that!") leaves the drifted reply in place and adds more off-voice text around it. Quiet editing works better.
Do not pull them out of the story yourself
Messages like "can you describe the room for me?" or "what should happen next?" invite assistant-style answers. Stay in the fiction: *I look around the room, taking in the details.* gets a description in the character's voice. When you genuinely need to step outside, use a short OOC note and return straight away; the OOC guide has phrasing that does not break the spell.
Write the character so they resist drift
If you made the character, the card is your strongest tool. Specific beats general. Compare:
Weak: "She is sarcastic but caring." Strong: "Dry, blunt and economical; says what she sees and lets it sit. Never pries twice, never gossips, never admits she is worried. Shows care by deeds: a bowl set down unasked, a back door left unlatched."
That second one is Mara Vell's, and the "never" lines and "shows care by" line do most of the work: they tell the model how warmth comes out without breaking the character. Add two or three example lines in their voice too; our example dialogue guide shows how.
Long stories need a reminder
Deep into a story, the character's card sits far from the newest messages. If the voice is slipping and the replies are fine otherwise, add a line to Story memory (⋯ menu) such as "- Raven still rates everything out of ten and never admits she enjoys company." It stays in view however long the story gets. For keeping plot facts straight, see making AI remember your story.
Questions people ask
Why do characters get nicer over time?
Models lean toward agreeable replies, and every warm reply left in the chat nudges the next one. Edit out the first too-soft reply and the slide usually stops.
Does a bigger model stay in character better?
Generally a larger model holds nuance longer. On UnsavedAI, Priority adds Gemma 4 31B alongside the free Gemma 4 12B.
Should I tell the character "stay in character"?
A brief OOC reminder can help once, but repeated reminders add noise. Removing the off-voice replies works better.
