创建

登录 ReadmeX

或

了解我们如何处理你的信息: 隐私政策

ReadmeX

陪你读懂 AI 动态

今天,想先了解什么?

选一份动态总结,或直接聊聊你的问题。

社区和关注动态总结只属于你自己。

资讯

开发者用15个角色化 Claude Code 会话开发应用

AI 总结

一名 Reddit 开发者称,其独自使用 15 个各负专责的 Claude Code 会话开发整个应用,职责涵盖 iOS、Android、Web、后端、设计、安全和运维等。方案使用独立目录、仓库任务卡、状态日志、Git worktree 和钩子;作者自述约七周内关闭了超过 2,600 张任务卡。相关工作流与产出数据均为个人自述,尚未经过独立验证。

深度见解

该帖提供了一套可操作的多会话方案,通过专业角色、任务跟踪、持久状态和隔离的 Git 工作区来组织 Claude Code。

One long Claude session kept getting worse. It filled up with unrelated context, forgot decisions and stepped on its own work.

So each area of my project now gets its own session, which I call an operator: iOS, Android, Web, Backend, Design, Security, Ops. Each opens its own folder, loads its own CLAUDE.md and works only there. Work that belongs to someone else becomes a card on a board in the repo. About 7 weeks in, that's 2,600+ cards closed.

The parts that matter:

  • A folder per operator, so each session loads only its job
  • A board in the repo: one markdown file per card, each naming who has the next step
  • A state log per operator, so I can clear a session anytime and lose nothing
  • A git worktree per operator, so sessions don't touch each other's files

What I learned:

  • Start with 2 or 3 operators. Parallel sessions burn through your plan fast.
  • Enforce rules with hooks. A rule in a file gets skipped. A hook that refuses the commit doesn't.
  • Keep instruction files short. Every line loads in every session.
  • Clear early. A fresh session with a good state log beats a long one that lost the thread.
  • Give every card one owner. Work assigned to everyone sits.
  • Make sessions check the code before saying "done" or "waiting on X".

#Claude Code#Anthropic#Git

原文约 2 分钟读完阅读原文
群聊

还没有评论,来说说你的看法。