给套件做版本对照再合并
Version and mix your harnesses
要点
- 单套件只比基线好约4%，六十五套合起来能解的任务远多于任一单独套件
- 新套件要对照旧版本跑评测，只有分数更好才合并，别凭手感改
- 别用第一轮提示成败评判；看它收到反馈后多快能学到能过关
- 把技能写成脚本而不是长说明；成功做过的活再收成可复用工作流
原帖开头
Any respectable engineer probably has some custom harness sitting on top of Codex or Claude Code. You keep adding instructions, tools, memory, skills, evals, sub-agents, hooks. Every few weeks you find something stupid the agent does and change the setup again. I've been readin …
