When the Code Writes Itself: Inside Yeti’s 2025 AI Hackathon

As a team of designers, developers and strategists, we believe staying ahead of the curve means constantly experimenting with emerging technologies. In March 2025, we put that belief to the test by hosting an internal AI Hackathon with a clear challenge: what can we build using only AI-powered development tools?

The rules were simple—and strict. Teams could only use AI development tools to build their projects. No hand-coded modules. No traditional dev workflows. Just prompts, APIs, and whatever magic (or chaos) the tools could deliver.

Our goal? To explore the limits of what’s possible with today’s AI development tools—pinpointing where they excel, where they fall short, and how they might help us deliver faster, more cost-effective solutions for our clients. To push those boundaries even further, one team was made up entirely of non-developers, tasked with seeing whether no-code AI tools could bring an app to life without writing a single line of code.

The Projects

Planning a better Disney day using no-code tools

Built entirely in Lovable with Supabase handling backend storage, our non-development team created one of the most surprising outcomes of the hackathon: a (partially) functioning Disneyland trip planner that allows users to select their must-see attractions, arrival time, preferred pace (chill or efficient), and whether they plan to park-hop. The app pulls real-time wait time data—scraped from Theme Park Wiki—to generate a smart, customized itinerary. The team pulled off live API integration, break scheduling, and dynamic planning logic without writing a single line of code. It was a powerful demonstration of just how far AI tools have come in enabling non-technical creators to build real, working products.
‍Read the Case Study!

‍Bop or Flop (Songlio Dupe)

Inspired by one of Yeti’s favorite (and frequently broken) games - Songlio - the team set out to build their own version: Bop or Flop. Using tools like CodeGuide, V0, Lovable, and Claude, they developed a multiplayer music guessing game where players join a room, choose songs from the Deezer API, and try to guess track titles from short audio clips. Supabase and WebSockets powered the real-time functionality, but syncing game states across multiple players proved to be one of the project’s biggest challenges—highlighting some of the architectural limitations of current AI tools. Despite those hurdles, the team successfully completed playable rounds, delivering a solid proof of concept—and a project they're excited to keep building.

Route Planning + Storytelling for Sulu

Our final project focused on adding intelligent route-building and AI-generated storytelling features to Sulu, Yeti’s custom tour creation platform (want to learn more about Sulu?). Using Cursor and Yurt, the team quickly spun up a web-based map editor where users could draw a tour route and export it as a KML file. From there, three sub-features emerged: a POI discovery tool that pulled in local landmarks and imagery via Mapbox and Wikipedia, an OpenAI-powered storyteller that generated location-specific narration based on user preferences, and an aggregated insights panel that pulled historical and cultural data about locations along the route. The only drawback? All of the code lived in a bloated, monolithic App.tsx file—highlighting the need for better modularity in AI-driven codebases.
‍Read the Case Study!

The Tools

Lovable
Lovable is a visual AI development platform designed to turn natural language prompts into full-stack web applications—without writing code. Loveable  integrates with tools like Supabase and GitHub, making it surprisingly powerful for more complex app behaviors. While it lacks real-time collaboration features, Lovable proved that with the right prompting, even those without coding experience can bring a working product to life.

Cursor
Cursor is an AI-enhanced coding environment that brings natural language assistance directly into your codebase. It allows developers to generate, refactor, debug, and explain code in real time using AI models—making it especially useful for developers familiar with a code editor but looking to speed up development or unblock tricky logic. At Yeti, Cursor became the go-to tool for most of our AI-driven development, however, Cursor does tend to generate monolithic code if left unchecked, which made modularity and architecture a key focus area.

Claude
Claude is a conversational AI developed by Anthropic, positioned as a helpful assistant for tasks requiring deeper logic and longer-form responses. During the hackathon, teams used Claude to work through more complex architectural problems, generate code snippets, and even refactor bloated sections of logic when other tools got stuck. While powerful, Claude’s token cost was notably higher than other tools, making it more suitable for high-impact tasks rather than frequent use.

V0
V0 is a design-to-code tool that transforms text prompts or visual mockups into working front-end React components. It excels at rapid prototyping and UI scaffolding, making it ideal for getting the look and layout of an app in place quickly. During our Bop or Flop project, V0 helped spin up early versions of the user interface, but it struggled to handle more complex application logic. It’s a great companion tool when paired with more logic-focused environments like Cursor or Claude.

CodeGuide
CodeGuide is a project planning tool that uses AI to help you generate detailed product documentation, technical architecture, and development specs before you write a single line of code. It’s particularly useful for scoping out ideas and aligning on technical vision. During the hackathon, CodeGuide helped teams clarify their app’s logic, define data models, and generate scoped prompts that could be used in other tools like Cursor or Claude. It’s best suited for teams that want to bring more structure and intentionality to AI development workflows.

Supabase
Supabase is an open-source Firebase alternative that provides authentication, real-time databases, and API support out of the box. While Supabase integrates well with many AI tools, it does require a bit of understanding to configure effectively, particularly for advanced use cases involving WebSockets or edge functions.

Mapbox
Mapbox is a powerful location and mapping platform used to display, edit, and interact with geographic data. In our Sulu route builder, Mapbox enabled users to draw, edit, and export custom tour routes with precision. It also powered place lookups and visualizations of points of interest. While robust and highly customizable, developers need to ensure that the correct variables and APIs are being used—something that AI occasionally struggled with.

OpenAI
OpenAI’s GPT-4 model served as a backbone for nearly every team during the hackathon. Whether powering narration scripts for the Sulu storyteller, assisting with prompt generation for Lovable, or debugging recursive logic, OpenAI provided reliable natural language support throughout. It was also used in some more inventive ways—like generating spoken-style tour scripts or simulating game logic behavior based on user actions.

Where AI Shines—and Where It Still Struggles

The Pros of AI Development Tools

Speed
AI tools allowed our teams to build functional applications—including front-end interfaces, backend logic, and third-party API integrations—in a matter of hours. What would normally take days or even weeks of manual development was accelerated dramatically. This speed was particularly evident in the Disneyland planner and Sulu route builder, where core functionality came together before lunch breaks were over.

Accessibility
Perhaps the most striking success came from the non-developer team who built a fully working app without writing code. Tools like Lovable made it possible for people with no engineering background to tap into the power of software creation. This accessibility has the potential to democratize app development and unlock new possibilities for teams who don’t traditionally have technical resources.

Creativity Boost
The low barrier to entry encouraged experimentation. Because there were no setup hurdles or risk of breaking existing codebases, teams felt more comfortable trying wild ideas, iterating rapidly, and exploring unexpected directions. The “Bop or Flop” team even generated mascots and game flowcharts on the fly, giving their project a level of polish that would’ve taken much longer with traditional processes.

Cost Efficiency
The cost savings for prototyping or MVP creation are significant. AI tools cut out a large chunk of manual work, allowing teams to test ideas quickly without incurring the high costs of full development sprints. This opens the door for clients with limited budgets to see tangible product demos early—and for Yeti to validate concepts faster than ever.

The Cons of AI Development Tools

Architectural Weakness
While AI tools are great at spinning up features fast, they often don’t enforce best practices around architecture. Several teams ended up with massive App.tsx files containing hundreds of lines of tangled logic. Without human oversight to break things into components or enforce modular structure, scalability and maintainability quickly became major concerns.

Collaboration Limitations
AI tools like Cursor and Lovable are primarily designed for solo use. During the hackathon, multiple teams struggled with syncing efforts, sharing credentials, or working in the same environment simultaneously. This lack of collaborative infrastructure made real-time teamwork difficult and created friction when handing off projects between teammates.

Debugging Complexity
When something went wrong with AI-generated code, it often failed in unpredictable ways. In some cases, it created recursive logic loops or redundant API calls that were hard to diagnose—especially without a clear understanding of the AI’s logic flow. Debugging required creative prompting, manual overrides, or reverting to earlier versions of the app entirely.

Geographic & Contextual Limitations
For location-based features—like those in the Sulu storytelling tool—AI often misunderstood geographic context or pulled in unrelated landmarks. Even with detailed prompts, it sometimes suggested content outside of the defined bounding box or hallucinated facts. This highlighted the need for tighter data constraints and more reliable source validation in future iterations.

Yeti’s 2025 AI Hackathon gave us a front-row seat to the possibilities and pitfalls of AI-assisted development. From no-code breakthroughs to full-stack multiplayer games, our teams pushed these tools to their limits—and discovered just how far they’ve come.

But we also learned some hard lessons along the way. In Part Two of this series, we’ll dive into the biggest takeaways from the hackathon: what worked, what didn’t, and how we’re planning to use those insights to shape future development processes.