The Dumbest-Looking AI Prompt Just Beat Months of Careful Game-Design Prompt Engineering
Claude Opus 5 was used to generate a playable first-person shooter from a very short prompt that asked for a Call of Duty–level game and told the model to split work into subagents and keep iterating against a harsh critic. The approach, called a “Gauntlet Loop,” used Claude Code features like isolated subagents, high-effort orchestration, and repeated fix-test cycles. The finished browser game reportedly used only generated assets, ran on Three.js/WebGL2, and contained about 55,000 lines across 11 subsystems, though its internal critic log still rated it below real Call of Duty. The same prompt was then reused by others, including James Altucher and another developer using OpenAI’s GPT-5.6 Ultra, while one developer achieved a similar result with a long, detailed prompt instead. The builds look impressive, but the claim that the model created the game “from scratch” is uncertain because shooter mechanics like mouse-look, WASD movement, and raycasting are heavily documented and likely well represented in training data.
