Ask HN：如何像 Anthropic 那样使用 Claude
1 分•作者: haint_•大约 1 个月前
我刚收到一封来自 Anthropic 的邮件。除了常规的新功能广告，其中有一段内容我觉得很有趣：
> 如何让 Claude 像 Anthropic 一样工作
>
> “我目前日常使用的模式是：两个首席代理互相监督，如果其中一个失败，另一个会重启它。这些代理会根据我同时进行的 8-10 个项目，委托给技术负责人或产品经理代理，每个项目又有 5-10 个独立代理，根据问题不同，这些代理可能是通才或专才。尽管如此，我每天的提示词输入量仍然只有 30-50 条，我的独立代理通常能自主工作 2-3 天。大约 60% 的交互对象是首席代理，35% 是项目负责人代理，只有 5% 的时候是出现问题需要介入。所有这些代理都直接与 SendMessage 工具通信。”
>
> – Daisy，Claude 代码工程师
这对我来说非常陌生，因为它和我使用 Claude 代码的方式截然不同。我目前的工作流程是：
* 对于新功能，我会创建一个会话，根据复杂程度，我会使用“计划模式”或“超级能力”等功能来集思广益、梳理需求，并制定计划/规范。这个过程通常占据了我与代理交互的大部分时间。
* 然后，在它实现功能后，我会打开一个新的会话，利用我的技能来审查 PR，并将发现的问题发布到 PR 上。
* 一旦有了 PR 的评论，我会将它们反馈给原始会话，并使用另一个匹配的技能来判断这些发现的问题并尝试修复。这个过程通常会来回进行大约 3 轮。
* 当审查会话满意后，我会要求原始会话执行一个涵盖其已实现所有内容的端到端测试计划。
我觉得我的工作流程自动化程度不够，验证环节仍然过于手动。但这样我也更安心，因为我阅读了每一份报告，并且对 PR 的质量有一个大致的了解，而无需阅读代码。
这里有人使用类似于 Claude 的工作流程吗，即利用专门的代理而不是通用的代理？你们能分享一下你们的经验吗？你们有可以让我阅读以了解更多的指南/博客吗？
查看原文
I just received an email from Anthropic. Besides the usual new feature ad, there is a piece that I find quite amusing:<p>> How to Claude like Anthropic<p>> "My daily driver currently looks like: two lead agents that keep each other accountable and restart the other if either fails. These delegate to tech lead or PM agents for the 8-10 projects I'm running at any one time, and each project has 5-10 IC agents, generalists or specialists depending on the problem. Across all of these I'm still only doing 30-50 prompts per day, and my IC agents typically work autonomously for 2-3 days. About 60% of my interaction is with the leads, 35% with a project lead, and 5% is when something has gone off the rails. All of these agents communicate directly with the SendMessage tool."<p>> – Daisy, Engineer on Claude Code<p>This sounds very foreign to me because it couldn't be more different from how I use Claude Code. My current workflow is:<p>- For a new feature, I create a session, and depending on the complexity, I will either use plan mode or something else like Superpowers to brainstorm, work out the requirements, and create a plan/spec. This process typically makes up most of my interactions with the agent.<p>- Then, after it implements the feature, I open a new session, use my own skill to review the PR, and post the findings on the PR.<p>- Once I have the PR comments, I give them back to the original session and use another matching skill to give a verdict on the findings and attempt to fix them. This process will typically happen back and forth for ~3 rounds.<p>- Once the reviewer session is happy, I ask the original session to perform an E2E test plan that covers everything it has implemented.<p>I feel like my workflow is not automated enough, and my verification loops are still too manual. But I also feel more reassured this way because I read every report and have a general sense of the quality of the PR without having to read the code.<p>Does anyone here use a similar workflow to Claude, that utilizes specialized agents instead of just the general-purpose one? Can you share your experience? Do you have a guide/blog I can read to learn more about it?