Now vibe coding, so learning hammer FE ?https://channel.meo-ai.comhttps://channel.meo-ai.com/posts/604https://channel.meo-ai.com/posts/604Mon, 20 Apr 2026 01:29:14 GMT<span class="post-title-source">《基于Andrej Karpathy观察的Claude Code行为优化指南》</span><br /><br />标签：<a href="/search/result?q=%23AI%E8%BE%85%E5%8A%A9%E7%BC%96%E7%A8%8B" title="#AI辅助编程">#AI辅助编程</a> <a href="/search/result?q=%23ClaudeCode" title="#ClaudeCode">#ClaudeCode</a> <a href="/search/result?q=%23LLM%E6%9C%80%E4%BD%B3%E5%AE%9E%E8%B7%B5" title="#LLM最佳实践">#LLM最佳实践</a> <a href="/search/result?q=%23%E4%BB%A3%E7%A0%81%E8%B4%A8%E9%87%8F" title="#代码质量">#代码质量</a><br /><br />总结：该项目将Andrej Karpathy对LLM编程缺陷的观察转化为可落地的CLAUDE.md规范文件，通过"编码前思考、极简优先、精准修改、目标驱动"四大原则，系统性解决AI助手常见的过度假设、过度工程化和无关修改等问题，帮助开发者获得更精准、简洁、可控的AI编程辅助体验。<br /><br />文章要点：<br />- 问题诊断：LLM常犯的错误包括擅自假设却不验证、过度复杂化代码、擅自修改无关代码等，Karpathy一针见血地指出了这些痛点<br />- 编码前思考原则：不确定时要主动提问而非猜测，有歧义时呈现多种解读，该拒绝时要敢于说"这样更简单"<br />- 极简优先原则：只做被明确要求的功能，不为单用场景造抽象，不把200行代码写成50行就算过关<br />- 精准修改原则：只碰该碰的代码，不动"看起来不顺眼"的邻居代码，自己的烂摊子自己收拾，但别碰别人留下的<br />- 目标驱动原则：把"加个验证"改成"写测试让非法输入失败，再让它通过"，给AI明确的验收标准，它会自己循环到达标<br />- 使用方式：支持Claude Code插件一键安装，或下载CLAUDE.md文件到项目根目录，Cursor用户也有对应规则文件可用<br />- 取舍提醒：这套规范偏向谨慎而非速度，简单改错别字不必上全套，但复杂任务能帮你避开返工噩梦<br /><br />文章URL：<a href="https://github.com/forrestchang/andrej-karpathy-skills" target="_blank" rel="noopener" title="https://github.com/forrestchang/andrej-karpathy-skills">https://github.com/forrestchang/andrej-karpathy-skills</a><a class="tgme_widget_message_link_preview" href="https://github.com/forrestchang/andrej-karpathy-skills" target="_blank" rel="noopener" title="A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls. - multica-ai/andrej-karpathy-skills"> <div class="link_preview_site_name accent_color">GitHub</div> <div class="link_preview_title">GitHub - multica-ai/andrej-karpathy-skills: A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's…</div> <div class="link_preview_description">A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls. - multica-ai/andrej-karpathy-skills</div> </a>