I'm shipping a bullet hell game on Steam using AI. Here's what nobody tells you about the gap between prototype and production.

A few months ago I released the demo for Codex Mortis, a necromancy-themed survivor/bullet hell, on Steam. It runs on a custom ECS engine in TypeScript + PixiJS + bitECS and was built almost entirely through AI-assisted development. Today it's live at Steam Fest, and Early Access launches March 19.

The first playable prototype took one prompt. Getting to production took two complete rewrites. That gap between "it works" and "it ships" is what I want to talk about.

The one-prompt trap

It started with: "Make me a Vampire Survivors-style game." I had a working prototype in minutes. Within hours of iterating I had sprites, abilities, synergies, companions. It validated my game idea in hours instead of weeks.

This is vibe coding. You describe what you want, AI builds it, you iterate fast. It's incredible for prototypes. But here's what nobody warns you about: vibe coding builds debt faster than it builds features.

After a few weeks of "add X," performance tanked, files grew to thousands of lines, and every new feature needed hacks on top of hacks. So I refactored. Proper game loop, separated rendering from physics. Then I tried 16x more enemies on screen and that broke everything again.

So I threw it all away and started over with a new stack, proper ECS, and batched rendering. This happened twice. And honestly it wasn't painful, because AI makes rewrites cheap. When rewriting takes hours instead of weeks, code stops being "your baby" and becomes a tool you swap out when it breaks.

Vibe coding vs. vibe engineering

Vibe coding is "make me X." You get fast results but zero architectural coherence.

Vibe engineering is "build system X using module Y, following pattern Z, respecting constraint W." You're still working verbally, but you're feeding AI architectural context about how systems connect, what already exists, and what constraints matter.

The shift is from just telling AI what you want to telling it what you want and how it fits into everything else. This matters because AI is a great programmer but a terrible architect. It writes excellent code to spec but it won't see the big picture, predict future needs, or maintain consistency unless you explicitly tell it to. Your prompts are your architecture.

The role shift nobody talks about

I have 10 years in gamedev as a solo dev, programmer, lead, and producer. I assumed my deep technical skills would be the main asset when working with AI. They weren't.

What mattered far more was my experience as a lead and coordinator. Working with AI in production means defining specs, reviewing output, catching architectural drift, running parallel workstreams, and making priority calls. That's not senior dev work, that's lead work.

On a good day I'd catch three bugs, spin up three Claude Code terminals in parallel, feed each one a problem with proper context, and ship three fixes simultaneously. The bottleneck was never writing code. It was managing the process. My role shifted from someone who writes code line by line to someone who defines what gets built and checks whether it actually makes sense.

If you're a lead or producer wondering whether AI makes your skills obsolete, it's the opposite. You're already trained for the job that AI development actually requires.

I shipped a TypeScript game and I still don't know TypeScript

Before Codex Mortis my TypeScript experience was zero. I'm a Unity/Unreal guy. Yet I built a production game with a custom engine in a language I'd never touched.

AI let me transfer universal knowledge about how engines work, ECS architecture, production pipelines, and how things break at scale into a completely unfamiliar environment. I never actually learned TypeScript. I knew what to build, and AI handled the how. The patterns and instincts came from me. The syntax came from AI.

This is the most underrated thing about AI-assisted dev: your domain expertise becomes portable. Ten years of gamedev knowledge didn't stay locked in C# or Blueprints. It became something I could deploy anywhere.

What this means for you

Start with vibe coding. Prototype fast. Validate if your idea is actually fun before you invest real time.

Know when to stop. When adding features requires more hacking than building, rewrite. AI makes it cheap enough that you shouldn't be afraid of starting over.

Transition to vibe engineering for production. Describe architecture, not just features. Give AI the context it needs to write code that fits into your system.

Stay the architect. AI executes. You decide what gets built, how it connects, and when to tear it down and start fresh.

Codex Mortis Demo is live at Steam Fest right now and hits Early Access on March 19. It started with one prompt and took two rewrites to get right. That's the real story of AI game development, and it's a lot less glamorous than "I typed one sentence and got a game."