跳到主内容
@wquguru
精选88VoxAI 产品与模型

Voxyz 分享 Claude Code 协调员技能,实现多模型并行开发

People are still running a row of Claude Code sessions and tabbing through them…

原文
发到 X
推荐理由

提供了具体的工作流架构与模型分工细节,结构清晰且包含可验证的数据,对开发者有较高的实操参考价值。

People are still running a row of Claude Code sessions and tabbing through them to babysit each one lol

人们仍然会运行一排 Claude Code 会话,并在它们之间切换标签页来逐一照看,哈哈。

So I'm sharing the coordinator skill I use all day, and all night while I sleep. Last time it rebuilt a product: ~𝟭𝟲𝟬 𝗣𝗥𝘀 𝗶𝗻 𝟯 𝗱𝗮𝘆𝘀, 107 task briefs, one session running the whole thing.

所以我分享我全天以及睡觉时都在使用的协调员技能(coordinator skill)。上次它重建了一个产品:3 天内完成了 ~160 个 PR,共 107 份任务简报,一个会话运行了整个流程。

npx skills@latest add https://copybook.page --skill coordinator --skill parallel-goal-fleet

Once it's installed, say this to Claude Code, Codex or the Claude desktop app: "You are the coordinator for this run. Goal: <what done means, and the check that proves it>. Use Herdr for the lanes. Run directory: ../<project>-fleet"

安装完成后,对 Claude Code、Codex 或 Claude 桌面应用说以下内容: "你是本次运行的协调员。目标:<完成的标准,以及证明完成的检查项>。使用 Herdr 进行车道管理。运行目录:../<project>-fleet"

Open Herdr in a side terminal and you can watch every worker and take over any of them.

在侧边终端中打开 Herdr,你就可以观察每个工作进程并接管其中任何一个。

Here's roughly how it runs: 1. Splits the tasks: one lane is one worker on one branch, working from its brief. Front end goes to Opus 5.5, back end and shipping to Codex Sol, PRs that touch money or login get reviewed by Codex Astra first, and Gemini 4 Argon can give a second opinion on design and copy 2. Opens a progress board: a terminal panel that refreshes every 20 seconds with who's doing what and who's handed back. A worker writes one READY line when it's done, and only that wakes the coordinator 3. Batch gates: ready PRs get tested together, and each batch ships with one push. Before batching, 𝗺𝗮𝗶𝗻 𝘁𝗼𝗼𝗸 𝟭𝟯𝟲 𝗽𝘂𝘀𝗵𝗲𝘀 𝗶𝗻 𝟯𝟮 𝗵𝗼𝘂𝗿𝘀, each one a paid build 4. Asks you: only shipping, database changes, anything touching money and deleting data make it stop and ask you. Once a ship worker saw the tests pass and went to push the payment PRs. Blocked on the spot 5. Overnight: it stops at the quota line you set, and when you wake up, the first thing you see is what's waiting on your yes

大致运行流程如下: 1. 拆分任务:一个车道对应一个分支上的一个工作进程,根据其简报工作。前端分配给 Opus 5.5,后端和发布分配给 Codex Sol,涉及金钱或登录的 PR 先由 Codex Astra 审查,Gemini 4 Argon 可为设计和文案提供第二意见。 2. 打开进度板:一个每 20 秒刷新一次的终端面板,显示谁在做什么以及谁已交还控制权。工作进程完成后写一行 READY,只有这行内容会唤醒协调员。 3. 批量门控:就绪的 PR 一起测试,每个批次通过一次推送完成发布。在批量处理前,main 分支在 32 小时内有 136 次推送,每次都是付费构建。 4. 询问你:仅发布、数据库变更、任何涉及金钱或删除数据的操作会使其停止并询问你。有一次,一个发布工作进程看到测试通过后准备推送支付相关的 PR,被当场阻止。 5. 夜间运行:它在设定的配额限制处停止,当你醒来时,首先看到的是等待你确认的内容。

Start by having it spin up one Claude worker and one Codex worker on a one-line task each. Firing up every worker at once is 𝘁𝗵𝗲 𝗳𝗮𝘀𝘁𝗲𝘀𝘁 𝘄𝗮𝘆 𝘁𝗼 𝗰𝗿𝘂𝘀𝗵 𝘆𝗼𝘂𝗿 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗮𝗻𝗱 𝘆𝗼𝘂𝗿 𝗾𝘂𝗼𝘁𝗮.

首先让它为每项单行任务启动一个 Claude 工作进程和一个 Codex 工作进程。同时启动所有工作进程是摧毁你的机器和你的配额的<span style="font-style:italic;">最快</span>方式。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件