跳到主内容
@wquguru
精选82VoxAI 产品与模型

Voxyz_ai分享Opus与GPT协作编程工作流

I updated my coordinator setup based on what everyone's been saying and my own t…

原文
发到 X
推荐理由

结构清晰拆解了多模型协作的具体分工与实战效果,有具体数据支撑,对同类创作者有参考价值。

I updated my coordinator setup based on what everyone's been saying and my own testing: Opus 5.5 as the coordinator + GPT-6.2 doing the work is 𝗮𝗻 𝗮𝗰𝘁𝘂𝗮𝗹 𝗴𝗼𝗹𝗱𝗺𝗶𝗻𝗲.

根据大家的建议和我的测试,我更新了协调器设置:以 Opus 5.5 作为协调器 + GPT-6.2 负责执行,这确实是一个真正的黄金组合。

The idea: every model only does the part it's best at. Better results, and your quota lasts longer, because every task lands on the right model.

核心理念是:每个模型只负责它最擅长的部分。这样能获得更好的结果,并且你的配额使用得更久,因为每项任务都能落到合适的模型上。

I tried running 6.2 Sol on Ultrafast as the coordinator. It's really fast, but it 𝗷𝘂𝗺𝗽𝘀 𝘁𝗼 𝗰𝗼𝗻𝗰𝗹𝘂𝘀𝗶𝗼𝗻𝘀 𝘄𝗮𝘆 𝘁𝗼𝗼 𝗲𝗮𝘀𝗶𝗹𝘆 and then just stops. Opus is still the better coordinator.

我曾尝试让 6.2 Sol 在 Ultrafast 模式下担任协调器。它确实很快,但会𝗷𝘂𝗺𝗽𝘀 𝘁𝗼 𝗰𝗼𝗻𝗰𝗹𝘂𝘀𝗶𝗼𝗻𝘀 𝘄𝗮𝘆 𝘁𝗼𝗼 𝗲𝗮𝘀𝗶𝗹𝘆(过早得出结论),然后就停止了。Opus 仍然是更好的协调器。

Here's the split, with reasoning effort: → Coordinator: Opus 5.5 xhigh. Splits the tasks, writes the briefs, checks the work, never writes a single line of code itself → Front end: Opus 5.5 high, xhigh for design drafts. Codex reviews, never edits → Back end: 6.2 Sol high, xhigh for the hard stuff and anything touching money. Every worker gets its own worktree and branch → Batch tests and shipping: 6.2 Sol medium. Takes a while, but it's all mechanical → Pre-merge review: 6.2 Astra xhigh, only on PRs that touch money, privacy or login

以下是分工安排,包含推理力度设置: → 协调器:Opus 5.5 xhigh。拆分任务、编写简报、检查工作,自身绝不写一行代码 → 前端:Opus 5.5 high,设计草稿用 xhigh。Codex 仅审查,不编辑 → 后端:6.2 Sol high,涉及核心逻辑和任何与金钱相关的部分用 xhigh。每个工作进程拥有独立的工作树和分支 → 批量测试与发布:6.2 Sol medium。耗时较长,但全是机械性工作 → 合并前审查:6.2 Astra xhigh,仅针对涉及金钱、隐私或登录的 PR

Last time we shipped payments, every test passed, and 𝗔𝘀𝘁𝗿𝗮 𝘀𝘁𝗶𝗹𝗹 𝗰𝗮𝘂𝗴𝗵𝘁 𝟯 𝗯𝘂𝗴𝘀 𝗿𝗶𝗴𝗵𝘁 𝗯𝗲𝗳𝗼𝗿𝗲 𝗺𝗲𝗿𝗴𝗲 that were about to go live.

上次我们发布支付功能时,所有测试均通过,而𝗔𝘀𝘁𝗿𝗮 𝘀𝘁𝗶𝗹𝗹 𝗰𝗮𝘂𝗴𝗵𝘁 𝟯 𝗯𝘂𝗴𝘀 𝗿𝗶𝗴𝗵𝘁 𝗯𝗲𝗳𝗼𝗿𝗲 𝗺𝗲𝗿𝗴𝗲(Astra 仍在合并前捕获了 3 个即将上线的 bug)。

Opus coordinates. Sol builds. Astra reviews.

Opus 负责协调,Sol 负责构建,Astra 负责审查。

Open Claude Code on Opus 5.5, send it the first prompt from the quoted post, and add this after it 👇

在 Opus 5.5 上打开 Claude Code,发送引用帖子中的第一个提示词,并在其后添加以下内容 👇

"Codex lanes run on gpt-6.2-sol. Only PRs that touch money or login get a gpt-6.2-astra review."

"Codex 通道运行于 gpt-6.2-sol。只有涉及金钱或登录的 PR 才会由 gpt-6.2-astra 进行审查。"

Already installed? Run npx skills@latest update. New to it: http://copybook.page/t/coordinator-skill-for-claude-code-and-codex

已安装?请运行 npx skills@latest update。初次使用请访问:http://copybook.page/t/coordinator-skill-for-claude-code-and-codex

Save it.

保存即可。

Oh and since this setup isn't tied to any harness, I've tried it in a bunch of apps. I like T3 Code a little more than the rest.

另外,由于此设置不依赖于任何特定框架,我在多个应用中试用过它。相比其他应用,我更喜欢 T3 Code。

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