Introduction

In Dec 2022, I built a Discord-native character generator using GPT-3 for my friends at Circle Labs. This was during AI Hack Week, a week-long hackathon hosted at the HF0 house. Since I was off work that week, I participated in the hackathon with Circle Labs and independently executed on a project we called “shape playground”.

what is circle labs?

Circle Labs is a venture-backed startup building a platform for kids to create AI-powered characters (called Shapes) that other kids can then interact with. They have ~100,000 shapes and an active, engaged community. You basically build a shape with your desired characteristics, chat with it individually on Discord, invite it to group chats, and deploy it on Twitter so it has a public presence.

what is shape playground?

Shape Playground is a character-generator I built from scratch for Circle Labs.

Previously, to generate shapes, shape creators submitted a blob of text describing the personality and backstory of the character, and someone on the Circle Labs team created the Shape for the shape creator, which would go live on Discord shortly after. This process had a few limitations:

1. there was no possibility for iterative development and testing of the character. When the shape creator sent the team this text, they had no idea how that text would impact how the character talked. An ideal UX would let the shape creator incrementally test how their character talked based on the current character description, and accordingly refine their wording.
2. the character description didn’t provide fine-tuning for the character’s behavior and speech, making the shapes uninteresting. If I asked different shapes a more general question, such as “how can I become a good cook?”, all of them responded with very similar content and style of speech. A better UX would give the shape creator more control over the character’s speech by exposing more parameters for the shape creator to play with.
3. the character generation and editing was very manual. Ideally, this would be automated so shape creator can directly create, edit and publish their shape.

Beyond just addressing the limitations, there was a more general greenfield opportunity ahead of us to create an innovative product and UX around character generation that shape creators love.

Product

I’m pretty happy with what I created here, since I believe it’s a novel approach to character generation.

It’s essentially a playground where you can create a character, test it, fine-tune it, and automatically deploy it when ready. It exposes a many useful knobs that help create rich characters. You can add an avatar, name, backstory, emojis, emotions, example responses, and more, all using generative AI, in a Discord-native, intuitive interface. I came up with the product concept through exploration and iteration, seeing what’s possible at the intersection of Discord api’s, prompt engineering, and conversational UX.

Below are two videos where I present a walkthrough of the product features.