How Astra 6 Rebuilt an AI-Made Browser Game

Summary

The article compares an original browser game made with ChatGPT 5 and some Gemini 3 with a later version, Reforged, built with Astra 6 through Hermes Agent. The original game used a single HTML file and centered on clicking an orb, buying upgrades, fighting bosses, and restarting with a prestige bonus. Reforged preserves that progression loop but adds a Three.js-rendered 3D tower, first-person walking with WASD and mouse look, interactive stations, upgradeable spells, target practice, a slot activity, and synthesized audio. Astra 6 wrote code for the scene, simulation, activities, interface, procedural Web Audio, and 3D ending, while delegated workers handled separate areas before integration. Human feedback led to a smaller HUD, clearer interaction prompts, visible spell trails and impacts, boss attack warnings, and Ward-block feedback. The article stresses that these improvements required repeated review rather than appearing correctly in the first generated version. Testing uncovered a balance flaw in which a player could finish the first run without using prestige, because the initial automated strategy kept buying upgrades and missed a no-purchase route. After a revised boss-health curve, 48 of 48 tested no-jackpot first-run strategies failed to finish, while tested prestige continuations still completed; the author notes this does not prove that no better strategy exists. A separate save-import bug allowed an extremely short boss attack interval to freeze combat, so accepted intervals were restricted to four to five seconds and regression tests were added. The final recorded verification passed 94 Node tests, 79 browser tests, the production build, and a check. The article explicitly says this is not a matched-prompt benchmark: the versions had different starting conditions, tools, briefs, and feedback, with no comparable time, token, or cost data. It presents the project as an example of AI-assisted game development that still depends on human direction, playtesting, and judgment.