用 Jev 构建代码库分类器
开发者分享用 Jev 构建代码库分类器,认为这可能解决智能体生成过度工程化代码的问题,并询问下一步测试方向。
作者在每次Claude Code消息中运行Jev(约0.3秒/次),用于判断值得记住的决策、bug等信息,并将结果保存到JEVMEM.md,比使用LLM(约3秒)更高效。
A 10th: deciding what a coding agent should remember. I run Jev on every Claude Code message (decision? bug? replaces an older one?) and save the keepers to JEVMEM.md in the repo. At 0.3s a call that's cheap enough for every message. An LLM would add about 3s each.