A tour inside my software restaurant · Nick Huber

How I vibe-coded my favorite defunct game in 7 days and most of my “users” ended up being bots.

For the past 18 months, I have had a private benchmark for coding agents: can they rebuild my favorite defunct game? (An autobattler called Storybook Brawl, which imploded after being acquired by FTX).

Every few months, I have tried again with the newest tools – in early 2025, with Cursor and Sonnet 3.7, in late 2025 with Claude Code and Opus 4.5, at some point I spent a weekend hacking on Lovable project with I think GPT-5, and more recently with a combination of both Codex and Claude Code.

Earlier attempts were impressive until they weren’t. I could get a basic template of something running, but complex interactions – the real meat of any strategy game – eventually became impossible. The agent would confidently tell me a change was finished when it either hadn’t worked or had broken three unrelated things. It was tantalizingly, agonizingly close – but no cigar.

This time, I can say that this benchmark is basically saturated. This is my journey.

FableBrawl: an overview #

In roughly seven days of full-time work, I built FableBrawl as a website and an iOS app. It has Google and Apple sign-in, subscriptions through Stripe1, a Vite-based game client, Postgres on Railway, deployment through GitHub Actions and Vercel, activation emails through Resend, analytics logs in Cloudflare R2, and a staging environment where I could test changes before shipping them. I had heard of most of these tools before but had never built on top of them; I can’t in full honesty say I still understand what all of them do, except to say that “it works.”

While some of the game engine was getting built in multi-hour refinements, it was nice to be able to do more emotive work like choosing visuals and sound effects in parallel. In the end, the game shipped with two visual modes: high-fantasy art from OpenAI image models, the original default aesthetic I personally wanted, and community-suggested MS Paint-style stick art for the anti-AI set (more on that later) that eventually became the default.

My background tracks for different game scenes were basically one-shot from the ElevenLabs music model and the sound effects come from their SFX model. Surprisingly, I had to generate ~30-40 sound effects for each that ended up in the game to get exactly what I wanted; the models still feel somewhat stunted when it comes to multi-modality and human taste – a lot of the SFX were really bothersome or overly shrill, at least to my ears.

Hear FableBrawl<br>Nothing plays until you tap an icon.

Main background track Music · 1:00<br>Play

Move a card Board interaction · 0:01<br>Play

Buy a card Shop interaction · 0:01<br>Play

Win a battle Victory cue · 0:02<br>Play

I made an admin dashboard to grant free subscriptions and monitor live usage. I built a comprehensive logging suite and in-app bug reporting feature. Every build ran about 400 automated tests, about a 100 I told the coding agent to write and about 300 I had it make for itself. I’m not a SWE so I don’t really know if these are good tests, but they seemed to catch some regressions.

Above all, I was able to faithfully and accurately resurrect the core game mechanics of my favorite defunct game with all of its crunchy, tactical decisions in an appropriate, high-fantasy setting – and be able to play on a live multiplayer server with other people!

It was especially fun to build suggestions from players in a few hours while they were still playing, like a Mortal Kombat-style “killing blow” cutscene that wasn’t even in the original game. Also, the in-game AI opponent (with no specific steering or customization – more on this later as well) was significantly better than what shipped with the original game.

A peek inside my software kitchen #

There is a heated debate on Twitter right now over if “software factories” – the idea of agents autonomously, continuously building high-quality software – are real or not. I’m not an expert, but what I will say is that my experience working with these agents felt more like running a slightly chaotic (but happy!) restaurant more than operating a factory.

Overall, my biggest takeaway is – damn, it’s fun to build ! The amount of iteration you’re able to do in just a few minutes or a single session made me always wanting to put more time into it, similar to that feeling of starting a new video game captured perfectly here:

At the same time, it was slightly chaotic because most of the time I was context-switching across ~20 threads to keep myself from being the rate-limiting factor towards progress, which ranged from making game assets to ruling on mechanic nuances to polishing the UI/UX to testing the game features. I enjoyed it when I had a few minutes to take a breath while waiting for “big dishes” to be ready, when I could more peacefully refine “would be nice” small touches that any small business owned and...