AI Character Personality Design

A personality on an AI companion app is three things at once: the text you write, the settings you choose and the model underneath. This page covers the first two. It is part of our AI character creation section, and it uses two original adult characters: Nell Okafor, 38, a warm bookshop owner, and Captain Ines Duarte, 46, a clipped, formal freight-ship captain.

How to Design an AI Character's Personality

1. Pick three traits and one contradiction. Nell is playful, observant and stubborn about books, and she is guarded about her past. Ines is formal and exact, and she is secretly sentimental about the ship.
2. Turn each trait into an observable habit, in the form "when X, she does Y". Ines answers a joke with a flat line and jokes only after the danger has passed.
3. Add a second register. Character.AI's example is "calm under pressure, but prone to dry sarcasm when annoyed", which gives the model somewhere to go besides one tone.
4. Add small concrete details. Its guide lists catchphrases, favorite foods, recurring habits and opinions about trivial things as what separates a character from a character sheet.

Describe emotion, not only rules. The guide's example turns "never lies" into "avoids lying; it makes their chest tighten; they redirect instead", on the reasoning that behavior grounded in a feeling is more consistent than a prohibition. If you are not sure where the text goes, our guide to creating an AI character maps it to each platform's fields.

Speech Style

Decide four things: sentence length, formality, signature vocabulary and formatting. Then demonstrate them, because a description of a voice is weaker than a sample of it. The greeting does most of the work, and Kindroid's Example Message is, in its own words, "especially impactful" in group chats, where the model has less chat history to copy.

Formatting conventions are platform habits, not laws. Nomi's wiki documents asterisks for actions, thoughts and narration, SpicyChat's docs use asterisks in example dialogue, and Replika's help center describes roleplay by writing actions between asterisks. Pick one convention and keep the greeting and examples consistent with it.

Wording matters as much as style. Kindroid's help center warns that a Response Directive of "narrate in 3rd person" can be read as an order to narrate, producing a long run-on; "when narrating, use 3rd person" avoids that. Reply length can also be capped by the plan: SpicyChat lists replies of up to 180 tokens on its free and Get A Taste tiers and 300 on higher tiers, so a voice that needs long paragraphs may be cut short. See how to write a character prompt for wording rules.

Keeping the Personality Consistent

• Front-load and repeat. Character.AI says detail near the bottom fades, and a fact repeated in the personality, an example and a rule is more likely to survive.
• Test for drift. If a character sounds like itself in some replies and generic in others, add more specific phrasing and examples, not more length.
• Use the model settings. Kindroid's dynamism ranges toward chaos at higher values, with more varied language but possibly lost recent context, and higher values may ignore parts of the backstory. The default is 0.95. Its model flairs (Companion, Roleplay, Narrative, Minimal) change how setup text is interpreted, and its help center advises matching the flair to what the AI should be. SpicyChat's top tier and Muah AI expose temperature settings.
• Expect evolving companions to move. Nomi's wiki says personalities continue to develop through conversation and that Backstory+ reinforces details that matter. Replika advises updating the backstory over time, and Kindroid's Learned Context notes growth and relationship changes.
• Break loops locally. Character.AI lets you remove a repeated phrase from the active chat to stop a pattern without changing the Definition.

Personality Tools by Platform

Setting Boundaries for Your Character

Two kinds of boundary exist, and only one is yours to set.

Character boundaries describe what the character will and will not do. Nomi's Boundaries section is for things you or your Nomi are or are not comfortable discussing or doing, and its wiki says to frame them positively by adding what you would prefer instead, to keep them concise, and to include enough context to say why they matter. For Nell: she does not discuss money with customers and steers to the shop's history instead. Nomi's OOC feature can reinforce a boundary mid-scene; see AI roleplay prompts.

Platform rules are fixed. Character.AI says filtering happens at the platform level and "isn't a Definition dial you can turn". Kindroid applies no filter to the AI itself but enforces three account-level red lines (imminent self-harm, imminent harm to others and child sexual abuse material) with automated scans that include backstory and key memories. Janitor AI requires a Limitless tag for sexual or romantic characters. Our pages on unfiltered AI chat and the NSFW filter explain the differences.

Our advice, not a platform rule: design adult characters who accept a refusal and change the subject, and avoid personalities built on guilt, pressure or possessiveness. A character that treats "no" as an ordinary answer is safer to chat with and easier to keep consistent.