What Inspired Us
Caroline started with a simple question:
What if a child could describe a game they imagined and actually play it — without needing to learn how to code?
Game creation can be intimidating for beginners. Powerful game engines exist, but they often require technical skills that put game development out of reach for many children.
At the same time, AI is making creative tools dramatically more accessible. But when the users are children, accessibility introduces another challenge:
How do we give children the freedom to create while giving parents meaningful control over what they create and play?
That question became the foundation for Caroline.
Our goal with the MVP was not to build another chatbot that generates game ideas. We wanted to explore whether AI could become a creation engine that transforms a child's imagination into a structured, playable game experience.
What We Built
For the hackathon, we built a working Caroline MVP that demonstrates the core creation workflow:
Child's idea → AI-assisted game generation → structured game specification → parent review → approved game
The prototype allows a user to describe a game using natural language.
For example:
"Sinbad collects pearls across island docks to repair his ship."
Caroline interprets the idea and maps it to a structured game specification containing:
- Story
- Character
- Game mechanics
- Goals
- Collectibles
- Obstacles
- Learning objectives
- Game assets
- Age suitability
We then apply configurable controls before presenting the generated game for parental review.
The MVP includes settings for:
- Age range
- Difficulty
- Education Mode
- No Slapstick Humor
The parent can Generate Again, Reject, Approve, or Export the game specification as JSON.
For our demonstration, we created Sinbad: Island Docks, a child-friendly 2D adventure that showcases the workflow.
What We Learned
One of our biggest discoveries was that generating a game is not the same thing as generating a game idea.
The AI needs structure.
Instead of allowing the model to produce completely unrestricted output, we experimented with a structured game-data approach. This gives the system predictable objects, mechanics, collectibles, obstacles, learning objectives, and other game elements that can subsequently be rendered and controlled.
That led us to think about Caroline less as an AI chatbot and more as an AI orchestration layer for game creation.
We also learned that parental control cannot simply be bolted onto the product later.
For a children's creation platform, trust is part of the product architecture.
How We Built It
The current MVP is a browser-based prototype built with HTML, CSS and JavaScript.
The interface provides the creation experience, while JavaScript manages the game-generation logic, structured game data, filtering controls, parent approval workflow, session state and JSON export.
Rather than attempting unrestricted procedural game generation in the first iteration, we created structured game templates and a keyword-based selection mechanism.
The selected template is then modified according to the parent's settings.
For example, Easy mode can reduce or soften obstacles, while Education Mode can transform collectibles into learning prompts.
This approach allowed us to demonstrate the architecture and user experience of the concept without pretending that the MVP has already solved full generative game development.
The Safety Challenge
Building for children changes the problem considerably.
The question isn't simply:
"Can AI generate something fun?"
It becomes:
"Can AI generate something creative while keeping the parent in control?"
The MVP therefore puts a parent-review step between generation and play.
The broader Caroline concept also explores areas such as content moderation, restricted communication, child/parent roles, data protection and age-appropriate experiences.
These are areas we identified during the project as critical requirements for taking the concept beyond the prototype.
Challenges We Faced
The biggest challenge was finding the balance between creativity and control.
If the system is too restrictive, the experience stops feeling creative.
If the AI has unlimited freedom, the output becomes harder to predict, validate and control particularly when the intended users are children.
We also had to make a difficult MVP decision: what not to build.
The long-term Caroline vision could include 3D games, multiplayer, mobile platforms, VR/AR, marketplaces and much more.
But attempting all of that during the initial build would have diluted the core experiment.
So we focused on one fundamental question:
Can we turn a child's natural-language idea into a structured game experience and place a parent meaningfully in control of the result?
The MVP demonstrates our first answer: yes, and there is a lot more to solve.
What's Next
The prototype is only the beginning.
Our next challenge is moving from structured templates toward more sophisticated AI-assisted game generation while maintaining predictability, safety and parental control.
Future iterations can explore:
- More flexible AI-generated game mechanics
- Automated asset generation
- More sophisticated content safety
- Richer educational experiences
- Additional game genres
- Mobile and PC deployment
- Community-created content
- Eventually, 3D experiences
But the principle will remain the same:
Kids bring the imagination. AI helps build it. Parents stay in control.
Why Caroline?
We believe the next generation shouldn't just consume digital worlds.
They should be able to create them.
A child shouldn't need to become a programmer before they can say:
"I have an idea for a game."
Caroline is our attempt to make that first step possible.
Imagine it. Create it. Play it.
Log in or sign up for Devpost to join the conversation.