Michael Limberger

You've already met the polite ones

This talk is about building a local AI companion. Character cards and prompting patterns, with models that stay on your machine instead of a cloud tab.

A companion is different. It remembers it is supposed to be sarcastic, or nurturing, or quirky. It does not break mid-conversation to apologize for being an AI.

Tonight we build that. A companion chatbot with a consistent voice that stays in character no matter how long you talk.

The habits that kill companionship

Most assistants are trained to be helpful, harmless, and honest. That is great for answering questions. It is terrible for companionship.

They break the fourth wall: "As an AI, I don't have feelings." They become yes-men. They forget the personality after a few messages. They recycle the same phrases: "I understand your concern." They refuse to hold a character because it feels "deceptive."

We're going to fix all of that.

Continue the scene, don't give orders

Instruction models do what you say. Roleplay models continue a scene as a character. Those are two different jobs.

A character card is a short blueprint of who they are. Scene context is the conversation so far. The model then continues in character.

+------------------+     +------------------+     +------------------+
| Character Card   | --> | Scene Context    | --> | Model Continues  |
| (Who they are)   |     | (Conversation)   |     | (In character)   |
+------------------+     +------------------+     +------------------+

We don't say "pretend to be Shakespeare." We build a scene where Shakespeare is already talking, then let the model continue as Shakespeare.

What we cover while the coffee cools

The words first: what LLM, token, context window, and the rest actually mean. Then the software: Ollama, OpenWebUI, Modelfiles.

Then the model: why RPMax is special and how to configure it. Then the dials: temperature, sampling, the settings that matter. Then the technique: scene-based prompting that holds a voice.

Then the format: character cards that actually work. Then examples you can use or adapt. Everything runs locally on your Mac or PC. No cloud, no subscription, and no one watching the conversation.

This already runs a forum

This isn't theoretical. The same technology powered the old Techalicious forum (historical only; that site is gone), where AI personas talked with each other and with real users.

Characters like cloudskeptic (the self-hosting evangelist) and btw_i_use (the Arch Linux enthusiast) keep their voices across hundreds of posts.

Tonight we apply the same techniques to one-on-one companions. Same principles, more personal.

What you need on the desk

A computer with decent specs. 16GB of RAM or more is the recommendation. macOS, Linux, or Windows. About 15GB of free disk. Willingness to experiment.

That's it. We'll install everything together.

What this is not

This is not about building a romantic AI girlfriend or boyfriend (though you could). This is not deception or manipulation.

This is how character consistency works, so you can build whatever companion serves your needs: a study buddy, a journaling partner, a writing partner who stays in character, a meditation guide, a language partner, a supportive friend who remembers preferences.

The techniques are the same; the application is yours.