Gendea

Inspiration

We were stuck with no ideas, and thinking harder wasn't helping. So we went meta: what if we built a tool for getting unstuck?

That's the problem Gendea solves. When you're stuck, when ideas won't come, or when everything you think of feels wrong, you need a way to break through. Not more thinking. Different thinking.

What it does

Gendea approaches ideation from three angles:

Spark takes a problem and returns multiple perspectives on it. You can draw a daily Tarot card to filter your session through a specific lens—a constraint that pushes your thinking somewhere unexpected.

Cauldron blends saved ideas together. You pick three, and it synthesizes them into something new.

Oracle is a Socratic conversation. It only asks questions, never gives answers. The goal is to reframe how you're thinking, not to solve the problem for you.

How we built it

We spent a lot of time on this. Initially we planned to submit multiple projects, but Gendea needed the work, so we focused on making it right.

We divided tasks between us and sometimes worked together on the same parts. LangChain was chosen specifically for structured JSON responses—prompting AI directly was unreliable. SQLite handles local development; PostgreSQL runs in production.

The day before submission, we swapped the Cauldron from 2D to 3D. The video shows the old version. We used specs for the rebuild.

Challenges

We originally built this for Frankenstein. We had an AI coach and a visual canvas, like Figma. Technically it was still about ideation, but there was too much going on. The UX suffered.

Cutting those features was hard. They felt useful. But removing them made everything easier to use. Once we trimmed it down, the project no longer fit Frankenstein, so we pivoted and leaned into the mystical theme instead.

Getting the UX right was the hardest part. We're still not sure it's perfect, but it's far better than where we started.

What we learned

We added local model support because it was free—we wanted to see how the app worked without burning API credits. Turns out local models can match cloud models, sometimes even beat them. That surprised me. I've been reaching for local models more often since.

Built With

• claude
• kiro
• langchain
• nuxt
• ollama
• postgresql
• spec-driven-development
• sqlite
• tresjs
• vue