Claude Code之父Boris Cherny近日在YC访谈中分享了关于AI产品迭代的新观点。[1]他建议开发者每六个月应当删除系统提示词、工具和harness代码，通过逐行删除并测试的方式，采用消融实验找到模型真正需要的指令。[1]这一建议源于Anthropic在7月24日发布的Claude 5上下文工程规则中，对Claude Code的系统提示词进行了大幅精简——删除比例超过80%。[1]
Boris认为模型能力往往远超现有产品的框架限制。[1]他提出了三条模型"解缚"的方法：给予模型更难的任务、进行更多实验，以及让模型能够自我验证成果。[1]为了验证这一理念，Boris进行了实践：他让Claude将一个Electron应用重写成Swift，该任务已经运行了14至15天，期间Claude主动建立了Slack频道并定期发送进度截图。[1]
Boris强调，编程学习者应当聚焦于应用能力而非纯理论知识，同时开发者需要放下对模型的控制欲，将AI模型视为同事而非工具来相处。[1]
Boris Cherny, the creator of Claude Code, shared unconventional views on AI product development during a Y Combinator interview, arguing that AI tools accumulate unnecessary constraints that limit model capabilities.[1] On July 24, Anthropic released updated context engineering guidelines for Claude 5 that removed over 80 percent of Claude Code's original instructions.[1] Cherny advocates a regular purge cycle, recommending that developers delete system prompts, tools, and harness code every six months through incremental ablation testing to identify which instructions the model genuinely requires.[1]
Cherny's philosophy centers on unleashing model potential rather than controlling it. He proposes three methods for removing constraints: assigning models increasingly difficult tasks, conducting more experiments, and enabling models to self-validate their outputs.[1] To illustrate this approach, Cherny described tasking Claude with rewriting an Electron application into Swift—a project that has been running for 14 to 15 days, during which Claude independently created a Slack channel to send progress screenshots regularly.[1] Cherny emphasizes that programming learners should focus on application capabilities rather than pure theory, and encourages treating AI models as colleagues rather than tools to be tightly controlled.[1]
