Detailed Analysis
A developer's weekend project has become a compelling case study in what advanced AI coding assistance can now produce: a fully browser-based, high-fidelity tribute to "Der Riese," the iconic zombies map from Call of Duty: World at War. Built with Anthropic's Opus 5 model running in "High" mode, the game—playable free at derkoloss.com—features solo and cooperative multiplayer with voice chat, a global leaderboard, and every visual and interactive element generated through code rather than pre-made assets. The map geometry, weapons, perk machines, mystery box, teleporters, 3D soldier models, particle effects, and even muzzle flashes were all procedurally created using Three.js, while ElevenLabs' voice and sound-effect generation tools supplied the audio layer. Within four days of release, the game logged over 1,000 sessions from players worldwide, and the developer has open-sourced the full codebase alongside raw multiplayer footage.
What makes this story notable isn't simply that a hobbyist with no formal game-development background built a working 3D multiplayer game—it's the granular, almost engineering-textbook detail the developer provides about the process. Rather than treating the AI as a one-shot generator, the creator spent 15-20 hours in iterative dialogue with Claude, refining the initial working prototype into something that could withstand real player traffic. The 12 documented techniques—many articulated by Claude itself when asked to summarize patterns from the session history—read like a practical playbook for AI-assisted software development. Insights such as "Claude fixes the exact thing named and stops" (requiring developers to generalize bug reports across entire classes of objects), or "give Claude eyes" by having it render outputs to images for visual comparison against reference art, reveal specific failure modes of large language models when handling spatial, physical, and perceptual tasks that have no ground truth beyond human judgment.
The deeper significance lies in what these tips expose about the current frontier of AI capability and its limits. Claude Opus 5 can apparently write "flawless netcode reconciliation" yet still place a crate blocking the only hallway in a level—illustrating a stark asymmetry between the model's logical/algorithmic reasoning and its physical or spatial intuition. The developer's solution—stating requirements as testable rules ("feet point at the floor when prone") rather than bug reports, then having Claude write validators that walk the map and prove the rule holds—represents a sophisticated methodology for compensating for AI blind spots. Similarly, the anecdote about a "measurement tool" reporting a false 0.0000 defect count due to a sign error underscores a recurring theme in AI-assisted engineering: verification code written by the same system being verified can inherit the same blind spots, necessitating adversarial testing (deliberately planting fake defects to confirm a tool can detect them).
This project fits into a broader trend of AI models moving from code-completion tools to genuine creative and technical collaborators capable of end-to-end asset generation—from 3D models to voice acting to sound design—collapsing traditionally siloed disciplines (art, audio engineering, gameplay programming, level design) into a single conversational workflow. It also reflects a maturing understanding within the developer community that state-of-the-art models like Opus 5 excel at velocity and one-shot prototyping but still require substantial human-guided iteration, systematic testing, and domain expertise to reach production quality. The reference to "Matt's gauntlet loop" methodology suggests an emerging shared vocabulary and best-practices culture forming around AI-native development, where prompting technique, verification design, and reference-driven iteration are becoming as important as traditional coding skill. As models continue to improve, cases like this—game development, complete with physics, netcode, and generative audio, executed almost entirely through conversation with an AI—serve as an early indicator of how creative technical work may increasingly be structured around human-AI feedback loops rather than manual implementation.
Read original article →