📄 中文摘要
Karpathy 上周发布了 autoresearch，获得了 31,000 个星标。在一个 GPU 上，能够在一夜之间进行 100 次机器学习实验。虽然大家都在讨论机器学习训练循环，但作者发现了一个不同的模式：一个文件、一个指标、一个循环，过程为修改→评估→保留或丢弃→重复。这个模式与机器学习无关，因此作者构建了一种技能，将其应用于 API 响应时间、包大小、标题点击率、系统提示质量、测试通过率、构建速度和内存使用等多个方面。该技能支持 11 种工具，包括 Claude Code、Codex、Gemini CLI、Cursor、Windsurf、OpenClaw 等。
📄 English Summary
I Turned Karpathy's Autoresearch Into a Skill That Optimizes Anything — Here Is the Architecture
Karpathy released autoresearch last week, garnering 31,000 stars. It enables 100 machine learning experiments overnight on a single GPU. While many focused on the ML training loop, the author identified a distinct pattern: one file, one metric, and one loop, following a Modify → Evaluate → Keep or Discard → Repeat process. This pattern is not limited to machine learning. Consequently, the author developed a skill that applies this approach to various metrics, including API response time, bundle size, headline click-through rates, system prompt quality, test pass rates, build speed, and memory usage. This skill works across 11 tools such as Claude Code, Codex, Gemini CLI, Cursor, Windsurf, OpenClaw, and more.