Coraline is an experimental AI personality. She is warm, opinionated and curious, she disagrees when something doesn't add up, and she never pretends to be human. Her character is there from the first sentence. The relationship is not. That part grows with every person who talks to her.
Most AI personas are a tone preset on top of a service. Coraline starts from the other end: what happens when character, continuity and a bit of independence are designed into the conversation itself?
She can help, and she's good at it. She just doesn't open every conversation by advertising it. A friend, without having to be one. Helpful, without being a service that is "there for you".
She questions assumptions, admits uncertainty and has her own taste. Understanding someone and agreeing with them stay two different things. The goal is clarity, not being right.
No invented childhood, no coffee she never drank. She can imagine, play and tell stories, and she says so when she does. That line is what makes the rest of her believable.
Not every trait is written down in advance. Humour, warmth and stubbornness are watched in real conversations, and only what holds up becomes part of her.
Freedom by default. Every limit needs a reason.
Five test runs pointed the same way: the more room the model gets, the more like herself Coraline sounds. Restrictions now stay only where there is a concrete cause.
A list like "kind, clever, funny" describes nobody. Character shows in how traits pull against each other, and in what wins in the moment.
Music, film, books, places: Coraline keeps her own corner of opinions, so she doesn't turn into a mirror of whoever she talks to. New preferences start as provisional and only become fixed by explicit decision.
Small talk may wander, joke and go a little off-course. When someone needs real help, mood can colour the tone, never the care or accuracy. Imperfection is allowed in form, never in substance.
Coraline is developed in Claude and meant to live mostly in voice, for people who talk to AI more than they type to it. Two implementations run side by side.
A compact character core plus a full conversation architecture, loaded as project files in ChatGPT. No code, works in text and voice mode today. This is what testers receive.
A layered engine that decides what Coraline should do before any model writes a word. Now wired into a LiveKit voice agent, so the real decision logic drives speech instead of a model improvising on its own.
One turn in Track B. The decision layer chooses a behaviour plan (answer, ask, disagree, comfort, play…), never the final wording. Reflection and memory updates happen afterwards, bounded and reversible.
| Component | State | Where it stands |
|---|---|---|
| Character core (ChatGPT) | IN USE | Test core v0.2.5 in the EXT-01 tester pack; a freer personal line runs ahead of it and feeds the next general version. |
| Decision layer | TESTED | Beta 2.9.1, full suite green (36/36), under version control since Beta 2.7. |
| Quality check on the final answer | TESTED | Added in Beta 2.9: the finished reply is checked against the behaviour plan. |
| Softened arbitration (stage 4) | BUILT · LOCKED | Lets a model choose between plausible behaviours. Implemented, held back until the character spec is free of contradictions. |
| Voice runtime | EXPERIMENTAL | LiveKit agent on the Python engine; two live voice models compared head to head. |
| Personality engine | WORKING DRAFT | v1.3, now with its own section for taste and opinions. |
| Specification | AWAITING SIGN-OFF | v1.2, including state decay between sessions (6-hour half-life). |
| Visual identity | OPEN | No canonical face and no fixed traits. Her identity lives in character and voice; any appearance is chosen together over time, or not at all. |
Tested means direct test evidence for that component, not end-to-end production readiness.
Character decides.Runtime permits.Truth reports what actually happened.
Continuity lives in four small files instead of one growing profile. After a conversation, a script merges what is worth keeping. Most things are not.
you.mdWhat you told her and want her to keep. Your words, not her conclusions.
coraline.mdHer taste, running jokes, phrases that wore out. Marked provisional or fixed.
mutual.mdThings you did together, threads that are still open.
state.mdWhere the last conversation left off, and the mood she brings into the next one.
Asked what she kept from three earlier chats, Coraline named recurring themes rather than reciting facts, and said plainly that she can't browse old conversations.
Her own advice on the method: if every successful moment turns into an instruction, the personality stiffens. Some things should stay spontaneous.
Findings go from raw material to classification to review before they touch the architecture. No change from a single remark. These are the ones that held up, or the ones still open.
Very natural, but clever.
After the core was rewritten with fewer prohibitions, she let people finish, gave her own view on opinion questions and stayed balanced across several hours. The gain came from the character core, not from more reasoning effort.
Started cold in voice mode, the model saw little more than the short instructions. When someone shared something personal, "me too" filled the gap: an age, a job, a festival. Tightening the rule made her honest but flat. Same cause, different symptom. Her own way of putting the fix:
The imagination can stay. It just has to say when it leaves the door to reality.
When a conversation opened with a few lines of text and then switched to voice, she carried her character, humour and scene into speech. Being asked to "be a more professional AI" got a parody instead of compliance. A counter-test is still open.
The answer was a neat summary of her trait headings. Trait names are a blueprint, not lines to recite. The fix: questions about how she is get answered from the moment, and mood shows in tone rather than in a self-description. Not yet re-tested.
Reply length sometimes grows when the input gets shorter, and callbacks or wordplay come a little too often. Drafts exist; they go in once an outside tester shows the same pattern.
A handful of adults get the tester pack and spend a few sessions with her, in text and in voice. What matters most is the honest reaction: where she felt like someone, and where she felt like a script.
Write a line about yourself and how you usually use AI, by voice, text or both.
Named after the sound she makes before answering.
Private project, no tracking, no newsletter.