ThreeJS Astronomically Accurate Air Show

While a developer might be seasoned and highly skilled, that doesn’t automatically make them an expert in every domain or tech stack. A brilliant iOS engineer isn’t instantly an Android engineer. A backend specialist isn’t automatically a game developer. Even if they can read and write the code, that doesn’t mean they can drive architectural decisions or move quickly in an unfamiliar ecosystem.

This is where an AI agent becomes transformative. We’ve already seen how powerful these tools are as enablers, someone with nothing more than an idea can now build a prototype or even a full app. But we’ve also seen the downside: “vibe‑coded” projects that technically work but are fragile, inconsistent, and difficult to maintain. The code often lacks structure, reliability, and readability.

Using AI as an extension of an engineer, however, is a different story. With direction, constraints, and supervision, an AI agent can produce work that is not only functional but elegant, efficient, and robust. And as LLMs improve, their built‑in technical competence improves too, meaning they can deliver higher‑quality results with less prompting.

In my own work, despite having built countless systems, I’m not a game developer. I don’t have deep experience with TSL, WGSL, shaders, or node‑based materials. But today, do I really need to? With an AI agent at my side, I can guide the process and still produce something substantial, performant, reliable, and visually striking.

For this project, I created a WebGPU‑based 3D projection of the real world. My AI agent wasn’t just a three.js expert, it was also a physicist and an astronomer. Together we built an environment with an astronomically accurate celestial system, day–night cycles, atmospheric refraction, a star field, a galactic disk, and a planet with real‑time weather systems and storm cells.

What would normally take weeks of research and development came together in a few hours at minimal cost. Through prompting, I was able to shape a sophisticated component that supports art direction, customisation, and performance constraints, all within a 200 kB bundle, with CPU usage under 10%, running smoothly across devices.

New engineering skill: AI‑assisted systems design

At least at the time of writing, AI is an accelerant for skilled engineers, not a replacement. But it does introduce a huge conceptual shift. I don’t need to memorise shader languages. I don’t need to know every API surface of three.js.

What I do need to know is how to specify constraints, evaluate output, and iterate. These are transferable skills, and arguably the ones worth learning from the very beginning.

AI‑generated vs AI‑supervised

I’ve observed that unsupervised AI code tends to be brittle and inconsistent. With supervision, however, AI can produce high‑quality, maintainable code, often beyond what either party could have produced alone. The engineer becomes a reviewer, architect, and director.

But this raises a real question: how comfortable should I be with only a casual understanding of how something works? How much “thinking” is safe to offload? It’s an ongoing tension, and one worth acknowledging.

A specialist at my fingertips

While I can describe concepts like star fields or weather systems, I certainly don’t know the names or mathematics behind orbital mechanics. I don’t know the algorithms that predict planetary positions, or how to translate those formulas into a 3D, animated, interactive environment that matches a specific visual style.

Atmospheric refraction normally requires understanding Rayleigh scattering and optical density gradients. I didn’t know any of that, but the AI did.

AI agents come with a wealth of general knowledge and domain‑specific expertise. Different models excel in different areas, intelligence, reasoning, performance, cost; so swapping models during development is common. The bottleneck shifts from knowledge to vision.

This globe became the perfect example of what’s now possible: a single engineer, amplified by an AI specialist, building something that once required a small team of experts.