用 Opus 5.5 做仪表盘动画并让模型自审画面
作者把质量检查从人脑搬进文件：CLAUDE.md 规则、check.py 数值校验、第二个模型当审稿人，Opus 5.5 从一张静图生成仪表盘动画并先读自己的帧再交人。
Everyone gets the same first result from a new model. It's good. So is the fifth one. And they're still sitting there, watching every render. Why? Because nothing else in the process can tell a good result from a bad one. The model stopped being the bottleneck. You are. You're the only quality check, and you don't scale. So I moved the check out of my head and into files: CLAUDE.md: rules read on every run check. py: numbers the model can't argue with a reviewer: a second model that didn't make the work /goal: keeps going until all of it passes Then it stops at a draft and waits for me. The dashboard in this video started as one still image. Opus 5.5 built it and read its own frames before I saw it. The full studio, file by file, in the article 👇
这条是引用着另一条写的：x.com/averix_official/status/2110711923968778244