I’m genuinely blown away. I’m a huge fan of Subnautica, and a few weeks ago I saw people creating games with the Gauntlet Loop. So, naturally, I asked Claude to build Claudenautica 🌊 I gave @claudeai Opus one prompt and ran Claude Code in Ultra mode. That’s it. No follow-up requests. No corrections. No additional guidance. It ran for 48 hours and consumed my entire weekly limit - and I’m on the Max plan 😅 After that, I let it continue working for a few more weeks until I felt the game was ready to try. During the entire process, I followed its progress through a live HTML page that the agent updated after every round - including its current work, status, comparisons, screenshots, and improvements. Today, I finally paused the loop and played the game for the first time. Here is the result. I’m open-sourcing the game and the exact prompt: 🎮 Play Claudenautica: github.com/sepivip/ClaudeNau… Here is the full prompt: “Build a playable underwater survival game at the visual and experiential level of Subnautica. It should feel like a real, beautiful alien ocean world, with everything at AAA quality - exploration, terrain, biomes, lighting, creatures, atmosphere, movement, survival, vehicles, bases, interiors, UI, effects, and anything else that matters. Break the goal into the smallest pieces that can be improved and judged independently. For every important piece, fan out a builder subagent and a separate harsh critic with fresh context. /loop on each piece. Critics must inspect the real playable/rendered output, directly compare it with representative real Subnautica gameplay screenshots and footage, blind A/B when possible, identify the biggest remaining gap, and send it back to be improved. Keep looping until our output wins against the Subnautica reference or I stop the run. Create a simple live HTML progress page in the project and continuously update it as you work, showing what is being built, current status, comparisons, and progress over time. Fan out subagents and ultracode.”