Intro

Me and my friends will plan a movie night and suggest movies before hand. But then it comes time for 10 of us to decide what movie we’re in the mood for.

The best we hand before this tool was going through the list and limiting the number of votes each person gets per round as we eliminate low scoring movies.

Tedious and time consuming are how I would describe this process.

This brewed in the back of my mind before I thought that I could use this as an opportunity to gain experience with AI coding tools to get this idea into the world.

In the start of the project I wanted the tool to be like a Jackbox game, reference the movies in my Jellyfin server, allow for a bracket competition between the movies to the death.

What AI I Used

I had experimented with Codex TUI with GPT 5.4, OpenCode TUI with Opus 4.6 and Kimi K2.5, then I tried Cursor’s Agent TUI with Composer 2. For me that was the sweet spot. Being a ‘price-conscious’ master’s student (aka broke), Composer 2 was great. I am happy with GPT 5.4’s performance in other side projects, and Opus seems really smart, but inefficient for my needs.

When I have an idea now I enjoy talking through my idea with Kimi K2.5 on T3.chat and fine-tuning my idea. Because I am experimenting with new tech and need to flesh out my idea before I take this idea into the terminal and start the planning stage.

In summary of this section, my process looks like this:

Flesh Out Idea (T3.chat) $\rightarrow$ Planning (TUI agent, Codex, or Cursor Agent) $\rightarrow$ Build (Same TUI agents) $\rightarrow$ Fix Issues (Same TUI agents, maybe trying different other agent or models)

What Tech was Used

I knew a couple things going into this project. I would use Vercel to host, for ease of use because I have my domain through Vercel. I originally wanted to experiment with Gleam to give functional programming a try, maybe with Lustre as the web framework. That was when I originally thought the project would be more minimal in scale and before I remembered about Convex and thought of how useful it would be for keeping clients up to date with each other. That was when I wanted to work with Convex for the database after rewatching Theo Browne’s video on it. Besides these initial opinions on what was used I gave no input on what the AI decided to use.

The model, Cursor’s Composer 2 mainly for this, decided to use my suggested Convex, Vercel along with: Next.js, TypeScript, Tailwind CSS v4, PostCSS, ESLint, and Turbopack. Which are technologies that I have never used, and knew little about beyond their general purpose and the bit from when I throw on a tech video to play in the background as I cook.

What I Learned

I am delighted when I get a product that is beyond my current abilities with a technology, for this project Next.js, Vercel, and Convex. But I am not really learning and more being impressed by the capabilities of these agents and models. This drove me to start another project in Odin, a language that I have never worked with, and Raylib, a library I have barely messed with in the C version. Now I try to work on this project of making a Ace Attorney style mystery game.

Now, I need to go think about how I can better use AI tools to learn the technology I want to use.

— Thanks for Reading,

Michael