Creating a real-time ray traced physics sandbox with Microsoft Copilot (www.hydrogen18.com)

A developer tested Microsoft Copilot by prompting it to generate complete C++ video games: first a working 2D “train defender” using SDL for I/O, Box2D for physics and ENTT for the ECS, then an attempted 3D first‑person shooter using SDL, OpenGL (and GLAD), Bullet Physics and ENTT. Copilot produced surprisingly large, nontrivial codebases — single-file C++ programs, shaders when asked, and even some rendering and damage simulation — but never produced a fully working 3D game. The author used careful prompt engineering (e.g., “output a single C++ file,” “do not generate a build system,” “include shaders,” “integrate Box2D with ENTT”) to steer results and discovered practical limits of the tool. Technically important takeaways: Copilot has output-size limits and will “drip” code across prompts; it hallucinates useful artifacts (invalid OpenGameArt URLs); it sometimes prefers verbose or obsolete implementations (OpenGL immediate mode) unless explicitly forbidden; and it often generates incomplete build/compilation steps. The experiment highlights Copilot’s strength as a rapid prototyping and learning aid but underscores that skilled developers are required to validate, refactor, secure and integrate generated code. For the AI/ML community this reinforces that prompt design, verification, dependency and API vetting, and security review remain essential when using LLMs to generate complex software.