Event ·Multi-source

Claude Code 新增 /build-eval 和 /hillclimb 命令,评测设计流程开源

Anthropic 的 RLanceMartin 在博客中把评测设计和爬山优化产品化为 Claude Code 的两个命令:/claude-api build-eval 和 /claude-api hillclimb,对应 Skills 开源在 claude-api skill 仓库。评测方面强调任务要镜像生产、评分器要先用程序化验证再用 LLM-as-judge,且裁判不能是被测模型。爬山方面核心是防过拟合:训练集升但测试集平就回滚,并禁止把失败案例内容直接粘贴进 prompt。两个案例给出了具体数字:客服成本优化从 74.4% 决策准确率、4.6 美分/单,降到 90.5% 准确率、约五分之一成本;claude-api skill 自身从 66% 提到约 88%。

Current read

Anthropic 把评测设计和爬山优化直接做成 Claude Code 命令了,还开源了 Skills,两条防过拟合纪律和客服降本 80% 的案例都挺实在。

11 sources61° AI heatLast updated 2026/9/29 02:36:47

Evidence

11 sources
Related source 1rohanpaul_ai
View source
Related source 2Gergely Orosz
View source
Related source 3Claude Code: GitHub Releases
View source
Related source 4Boris Cherny
View source
Related source 5claudedevs
View source
Related source 6Latent.Space
View source
Related source 7Latent Space (swyx)
View source
Related source 8AI产品黄叔
View source
Related source 10Tibor Blaho
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.