Key facts
- Claude Opus 5 built a playable first-person shooter game using a simple prompt.
- The AI model's output was described as 'utterly perfect' by the prompt's creator.
- The game was constructed entirely from custom code, with no external assets.
- The 'Gauntlet Loop' method involves setting a high bar, using subagents, and a critic.
- Other AI models and developers have replicated the prompt with successful results.
- Concerns about data contamination have been raised regarding AI-generated code.
Claude Opus 5, an advanced AI model, has demonstrated the ability to construct a fully playable first-person shooter game with a remarkably simple prompt, challenging conventional approaches to AI-driven game development. The prompt, authored by AI investor Matt Shumer, instructed the model to create a game comparable to recent Call of Duty titles, utilizing a method he calls the 'Gauntlet Loop'. This approach involves dividing the task among specialized subagents and employing a dedicated critic agent to evaluate and refine the work, rather than relying on detailed, prescriptive instructions.
Shumer's prompt, consisting of only three short paragraphs, eschewed the typical advice of specifying criteria for 'good' code. Instead, it tasked the AI with aiming for 'utterly perfect' results and delegated the definition of quality to a critic agent built by Opus 5 itself. This method contrasts with traditional prompt engineering, which emphasizes detailed specifications. The resulting game, built entirely with custom code and running on Three.js and WebGL2, features approximately 55,000 lines of code across 11 subsystems, with all assets generated dynamically.
The effectiveness of this approach has been validated by other developers. James Altucher reportedly used the same prompt to create a browser-based game called 'Operation Blackout' in about ten hours, consuming 1.3 million tokens. A developer also tested a similar prompt on OpenAI's GPT-5.6 model. In contrast, another developer, Leon Lin, created a game called 'Dust Corridor' by reverse-engineering a highly detailed, 20-section prompt for Claude Opus 5, highlighting the divergent strategies in AI game development.
However, questions have been raised about the extent of the AI's originality. Researchers point to the issue of 'data contamination,' where models perform well on tasks because similar code already exists in their training data. While Shumer's 'Gauntlet Loop' is recognized as a genuine advancement in agentic coding, the claim of building a game entirely 'from scratch' is tempered by the possibility that the AI recombined existing code patterns from well-documented genres like first-person shooters.