Felix Rieseberg says Claude workflows need raw data, tools and adversarial review
In an interview republished by 36Kr, Anthropic Claude engineering lead Felix Rieseberg described rebuilding his team’s workflow around agents with access to original data, tools and sandboxes rather than using Claude only to summarize or rewrite text. His team built Claude Cowork in roughly a week and a half, and he said he now focuses on generating prototype variants and using separate Claude sessions to challenge and judge the results. He also argued that agents should receive broad data access through documented tools instead of having small, preassigned datasets.
The report offers a concrete alternative to using Claude only for summarization: connect it to original data and professional tools, then check its output through adversarial review.
在最近一场深度访谈中，Anthropic Claude工程负责人Felix Rieseberg分享了他如今的工作方式，以及为什么大多数人仍然在用2024年的方式使用Claude。他的团队用大约一周半时间搭建了Claude Cowork，而他现在几乎不再写代码，自称“只是Claude的一条传送带”。核心观点可以概括为：如果你只是让Claude总结邮件、润色文字，你并没有真正从新智能中受益；真正有价值的是重新设计工作流程，让AI接触原始数据和工具，用它做原型、做对抗性审查、做你以前做不到的事；而人类最重要的能力，仍然是同理心、品味和判断力。
以下为 Felix 的完整分享：
一、我已经不写代码了，我只是 Claude 的传送带
我现在不再写代码。大多数时候，我在做原型设计：我们尝试很多想法，睡两晚，判断这是好主意还是坏主意，然后告诉 Claude 把所有变体做出来，再批判性分析哪些想法好、哪些不好。