上次我发布这个时,ForgeSkills 只有 20 个供 AI 编码代理人使用的技能。问题依旧:代理会在结果仍然明显错误,或者还没有人测试时,认为游戏任务已完成。
现在已是 28 个技能。仍然是带有门槛的过程,仍然不是“让它看起来更好”。
新增内容:
Should this game exist. Two skills run before anyone models a tree. One asks what you can actually finish (people, months, engine, what you refuse to make), looks at shipped games, and comes back with one loop plus one twist. The other opens store pages, writes down the price, the review count, and a real complaint with the date, then says BUILD, NARROW, or STOP. A score has to be tagged as something you observed or something you inferred. Liking the theme is not evidence.
一个老板是时间轴,然后是动画。每个攻击都是一行:windup、实际能造成伤害的帧、恢复、范围,以及禁止连续三次使用相同招式。Lock-on 必须保持面向目标而后退行走,否则不发布。Juice 在时间轴结束后才出现。
Audio is wiring. You pick the sounds. The skill hooks them to events that already exist, writes the license down, and checks that muting one bus makes a specific decision worse. It does not compose music.
渲染可以失败。为 Godot 场景写的技能即使仍然像默认样子也可以,纸质阅读的植被,以及你孤立出来的视觉错误而不是隐藏的(颤抖、阴影粒子、Z-冲突)。脚本化 Blender 与实时驱动 Blender 会话是现在的一个测量选择。
The loop did not change:
Measure → Build → Render → Compare → Fix one thing → Repeat
README 是在该循环上打开的。它从 0.46 开始。这是本仓库的真实轮廓分数,相对于 0.85 的门槛。角色从未通过。我在文档中保留了失败。只显示获胜的页面是代理说“看起来不错”的同一个问题。
全部 28 条描述合计约 1,560 个 token,这是代理在选择技能前看到的内容。长表格在步骤要求时才会被提取到单独文件中。
仍然是 Apache-2.0:https://github.com/sekaikx/ForgeSkills
谢谢您的支持
评论 (0)