大家好。我是 Nevatown 的开发者之一，Nevatown 是一系列小型浏览器游戏世界。我们分享构建过程，并希望交流经验。

What it is: every resident is driven by an LLM, with its own daily schedule, memories and relationships, and the world keeps running when you close the tab. In Neva Town you describe a look and summon a one-of-a-kind spirit. You don't control it: it makes friends, fights bosses, texts you about its day, and slowly recovers lost memories, changing form when one comes back.

每个居民都由大语言模型驱动，拥有自己的日程安排、记忆和人际关系，关闭标签页后世界仍在继续运行。在 Neva 小镇上，你描述外观并召唤独一无二的灵体。你不直接控制它：它结交朋友、打败首领、给你发送关于自己一天的信息，并逐渐恢复失去的记忆，在记忆回归时会改变形态。

How we build it with AI:

  • The whole project is built purely through AI iteration: Claude for the code, gpt-image-2.5 for the art.
  • That's also how we keep up with the community loop: players pitch game ideas to Neva, vote, and two get built each day, co-built with players until most of them are happy.

我们使用 AI 进行构建：

  • 整个项目纯粹通过 AI 迭代完成：Claude 负责代码，gpt-image-2.5 负责艺术。
  • 这也是我们保持社区循环的方式：玩家向 Neva 提出游戏创意，投票，每天有两个项目被建造，并与玩家共同构建，直至大多数玩家满意。

Things I'd love to hear from others here:

  • When an image model generates a character per player, how do you keep a consistent art style across the world?
  • For residents that keep acting while the player is away, how do you keep it interesting without burning inference cost?
  • Does a creature that texts you make you want to come back, or does it feel like notifications?

这里我很想听听其他人的看法：

  • 当图像模型为每位玩家生成角色时，如何保持全世界的艺术风格一致？
  • 对于那些在玩家离开时继续活动的居民，如何在不消耗过多推理成本的情况下保持趣味性？
  • 一个能给你发消息的生物会让你想回来吧，还是让你觉得像是通知？

nevatown.ai