Jonathan Bossenger

A little over a year ago, I had an idea. I wanted to test just how far AI-assisted coding had come by giving it a genuinely ambitious task: build a complete 3D space shooter game from scratch, guided only by a game design document and a set of requirements, in as few prompts as possible.

The project was called Cosmic Defender 3D. The result, in March 2025, using Cursor and the current Anthropic model, was a complete mess.

Fast forward to February 2026. I tried again — this time using Claude Code with the Opus model — and the result was a fully working game.

Same concept. Same ambition. Wildly different outcome. So what changed?

The Original Experiment (March 2025)

The idea behind the experiment was straightforward: write a thorough game design document and a requirements spec, then hand them to an AI and see if it could build the whole thing with minimal hand-holding. I wanted to simulate what a developer might actually do — give the AI context and intent, then let it run.

The 2025 attempt failed in almost every meaningful way. The AI could generate code — plenty of it — but it couldn’t hold the bigger picture together across multiple steps. Each prompt was essentially a fresh start. Context fell apart. The 3D rendering was broken. Game logic was inconsistent. What should have been a cohesive experience was a pile of disconnected pieces that didn’t quite fit together. If you visit the original deployed version, you’ll see exactly what I mean: it barely runs, and what does run doesn’t make much sense.

At the time, I wrote it off as a limitation of the medium. AI coding assistants in early 2025 were impressive in short bursts — autocomplete on steroids, essentially — but they weren’t truly agentic. They couldn’t plan, execute, verify, and iterate the way a developer would. They could write a function, but they couldn’t build a system.

The 2026 Rebuild

Today, inspired by some sharing some of my early failures with work colleagues, I tried a rebuild. I used the same game design document and requirements I had written for the original attempt. I fed them into Claude Code, running the Opus model, and essentially said: build this.

I found one bug during testing; the keyboard movement was affected by the mouse controls. I fed that data back to Claude.

The end result is a genuinely playable 3D space shooter. It’s not a AAA title, but it’s a complete, functional game — built from a design doc, in a fraction of the time the original failed attempt took, with far less frustration.

What This Means for Developers

I want to be careful not to overstate this. Agentic AI coding is not magic, and it won’t replace developers. The Cosmic Defender rebuild still needed a human who understood what a good game design document looks like, who could evaluate the output, and who knew enough to intervene when needed. The AI is a capable, tireless collaborator — not an autonomous replacement for engineering judgment.

But the gap between “AI as autocomplete” and “AI as collaborator” closed dramatically in the past year. Tasks that were theoretically possible in 2025 but practically painful — building a non-trivial project from a specification, refactoring a large codebase, implementing a feature end-to-end — are now genuinely approachable with agentic tools.

For developers willing to invest in learning how to work with these tools effectively — writing better specs, structuring projects to play to AI strengths, reviewing and guiding output rather than fighting it — the productivity gains are real and significant.

A Year Is a Long Time

What strikes me most about this experiment is just how much changed in roughly twelve months. The 2025 version of this attempt wasn’t a near miss — it was a complete and catastrophic failure. The 2026 version works. That’s not incremental progress; that’s a step change.

If the next twelve months bring a similar leap, the landscape for software development is going to look very different by early 2027. I’m not sure exactly what that looks like yet, but the Cosmic Defender experiment gave me a pretty clear signal that we’re moving faster than most people realize.

If you want to dig into the project yourself, the repos are on GitHub. The original broken version is still live at https://jonathanbossenger.github.io/cosmic-defender-3d-0.1/, with the source code at https://github.com/jonathanbossenger/cosmic-defender-3d-0.1, and the working rebuild is at https://jonathanbossenger.github.io/cosmic-defender-3d-1.0/, with the source code at https://github.com/jonathanbossenger/cosmic-defender-3d-1.0.