我一直在尝试使用 AI 辅助游戏开发，最让我困扰的一点是 AI 代理人可以轻易地宣称任务完成，即使结果仍然明显错误或功能实际上没有被测试。

于是我开始构建 ForgeSkills，一个包含 20 个可重用技能的开源项目，专注于真实的游戏开发工作流，而不仅仅是生成代码。

它的理念是让代理人遵循具体的操作步骤，而不是依赖模糊的指令，如“让它看起来更好”或“修正动画”。

Some of what's included:

  • Reference-to-3D reconstruction: Rebuilding assets from reference images and tracking visual differences instead of making something vaguely similar.
    参考图到 3D 重建： 从参考图像重建资产并跟踪视觉差异，而不是构建看起来大致相似的东西。

  • Blender automation and procedural modeling: Creating and iterating on game-ready assets.
    Blender 自动化与程序化建模： 创建并迭代游戏化的资产。

  • Animation and locomotion: Addressing things like foot sliding, awkward transitions, and movement that feels wrong.
    动画与角色移动： 解决脚滑、尴尬的过渡以及移动感觉不自然的问题。

  • VFX and game feel: Impact effects, hit stop, camera response, and coordinated feedback.
    VFX 与游戏手感： 冲击效果、击中停顿、相机响应以及协同反馈。

  • Gameplay engineering: Implementing game systems with explicit requirements and validation.
    游戏玩法工程： 在明确的需求和验证下实现游戏系统。

  • Debugging and verification: Reproducing issues, isolating causes, testing fixes, and checking that the result actually works.
    调试与验证： 复现问题、定位原因、测试修复并确认结果确实可行。

The core workflow is:

Measure → Build → Render → Compare → Fix → Repeat

核心工作流是：

Measure → Build → Render → Compare → Fix → Repeat

There are also supporting Python tools for parts of the workflow that can be measured or checked automatically.

此外，还有用于工作流中可测量或自动检查的 Python 支持工具。

The repository currently includes Godot and Unity pipeline skills, alongside Blender and engine-independent workflows. It's designed to work with AI coding agents that support the standard SKILL.md format.

仓库目前包含 Godot 和 Unity 流水线技能，以及 Blender 和与引擎无关的工作流。它设计用于与支持标准 SKILL.md 格式的 AI 编码代理配合使用。

GitHub: https://github.com/sekaikx/ForgeSkills

它是开源的，遵循 Apache-2.0 许可证。

This is still a work in progress, and I'd appreciate feedback from people who use AI agents to build actual games. I'm particularly interested in what breaks in practice, which workflows are missing, and how to make the skills more reliable with weaker models.

这仍在进行中，我很期待来自使用 AI 代理构建真实游戏的人的反馈。我尤其关注实际中出现的问题、缺失的工作流以及如何让弱模型下的技能更可靠。

If you try any of them, I'd love to hear what worked and what didn't.
如果你尝试其中的任何一个，我很想听听它工作得如何以及哪里出了问题。