用廉价助手砍九成花费
Cheap workers cut 90% tokens
要点
- 多数助手活是开文件和套模板，几乎不需要高价模型判断
- 廉价助手读文件出摘要、按样例写重复代码，贵模型不看原文
- 提示里的规则会被无视；超过350行要硬拦截再改走廉价通道
- 真正改文件仍需原文；廉价助手会漏掉贵模型几秒能抓住的bug
原帖开头
Spotify just published the internal setup their engineers use with Claude Code, the one that cut token usage by 90 percent. Here is how it works. The observation behind it: most of what a coding assistant does is not thinking. It opens five files to answer a question about one o …
