全自动化人工智能工作流程
我刚刚构建了一个系统，在预推送时，通过命令行监控对 GitHub Actions 的更改，并将错误反馈给 Claude。我在五个不同的编辑器中克隆了这个仓库，并设置了一个提示，使其能够始终开发、提交和推送有效的代码。这个系统在很大程度上依赖于静态分析和使用 Biome 进行代码检查，以确保最大程度的正确性和强有力的测试指导。
这个仓库是一个单一代码库，包含可以独立开发的紧密相关的服务。采用主干开发，所有更改直接推送到生产环境（部署后会进行一次烟雾测试检查）。
我不确定是否还有其他人构建过类似的东西，但本质上，我可以在不同的文件夹中“git clone”一个新版本的仓库，通过花几分钟创建一个提示，实际上可以实现自我复制。
这个工作的99%都是设计提示和调整各个部分，使其能够独立运作，以便我能够适应上下文窗口的最近部分。
这到底是怎么回事？
查看原文
So I just built a system that on pre-push, watch changes to GH actions using CLI and feed errors to Claude. I am cloning the repo in 5 separate editors with a prompt that makes it develop, commit and push working code at all times. It leans heavily on static analysis and linting with Biome to ensure maximum correctness and strong guidelines for testing.<p>Repository is a monorepo which has cohesive services that can be developed independently. Trunk-Based development and changes go straight to production (there's a post-deploy check that runs smoke tests)<p>I'm not sure if anyone else has built something like this, but essentially I can "git clone" a new version of the repository in a different folder and literally multiply myself by just spending a couple minutes creating a prompt<p>99% of the Job is designing prompts and adjusting the pieces to be independent so I can fit in the recent part of the context window<p>WHAT THE ACTUAL FUCK?