Claude Code 工作流:利用 Opus 5.5 与 Sol 双模型协作
Claude Code tip: if Opus 5.5 is your orchestrator, don't let your ChatGPT quota…
这篇内容详细拆解了如何利用不同 AI 模型的特性(如 Sol 的规划能力与 Opus 的统筹能力)构建严谨的代码审查闭环,结构清晰且提供了可直接复现的配置步骤,对开发者优化 AI 辅助编程流程极具参考价值。
Claude Code tip: if Opus 5.5 is your orchestrator, don't let your ChatGPT quota sit idle. Before a big change, have Sol in Codex write its own plan 𝗯𝗲𝗵𝗶𝗻𝗱 𝗢𝗽𝘂𝘀'𝘀 𝗯𝗮𝗰𝗸.
Claude Code 技巧:如果你的编排器是 Opus 5.5,别让你的 ChatGPT 配额闲置。在进行重大更改之前,让 Sol in Codex 在 Opus 的背后制定它自己的计划。
Claude Code can run codex exec straight from the terminal, which is like bringing in 𝗮 𝘀𝗲𝗰𝗼𝗻𝗱 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁. Sol reads the code, tries things out, and writes its plan in its own worktree, where it can't touch your working tree. Opus 5.5 stays in charge, and big changes go through three steps: → Before picking a plan: each writes one, and Sol never sees Opus's → Once both are in: Opus lists where they disagree, asks which way you lean, and builds only after you decide → After the code is written: a fresh reviewer writes tests from the requirements first, then reads the code and runs them Sol only plans. Writing and reviewing the code stay in Claude Code.
Claude Code 可以直接从终端运行 codex exec,这就像引入了第二位架构师。Sol 阅读代码、尝试各种方案,并在其自己的工作区中编写计划,在那里它无法触及你的工作树。Opus 5.5 保持掌控,重大更改分三步进行: → 在选择计划之前:各自编写一份计划,且 Sol 看不到 Opus 的计划 → 当两者都就绪后:Opus 列出分歧点,询问你倾向于哪一方,并在你做出决定后才开始构建 → 代码编写完成后:一位新的审查者首先根据需求编写测试,然后阅读代码并运行这些测试 Sol 仅负责规划。代码的编写和审查仍由 Claude Code 完成。
The tree is the full setup. Hand the tree and this prompt to Claude Code to set it up 👇
该目录结构即为完整配置。将此目录结构和以下提示词交给 Claude Code 以完成设置 👇
"Set up my Claude Code to match this tree: 1. Run codex exec -m gpt-6.1-sol --skip-git-repo-check "Reply with OK" once to confirm you can call Sol. If it fails, stop and tell me. 2. Add a rule to ~/.claude/CLAUDE.md: for architecture decisions or big changes, first start Sol in the background (codex exec --worktree -m gpt-6.1-sol -c model_reasoning_effort=high -s workspace-write -o /tmp/sol-plan.md) and have it write its own plan. Give it only my original request and let it decide how to work. If there are relevant uncommitted changes, pass it the diff too. Meanwhile, write your own plan. When both are ready, list where they disagree and what you recommend, ask which way I lean, and wait for my call before building. 3. Create a reviewer subagent in ~/.claude/agents/. Give it all tools, and don't add isolation: worktree (it wouldn't see uncommitted changes). Hand it my request and this change's diff. It decides how to verify, ideally writing tests from the requirements before reading the implementation, but it has to actually run them. It can write tests but doesn't change the application code, and it reports what it found along with the tests it wrote. Then add to ~/.claude/CLAUDE.md: before you report a big change as done, have the reviewer check it.
"按照此目录结构配置我的 Claude Code: 1. 执行一次 codex exec -m gpt-6.1-sol --skip-git-repo-check "Reply with OK" 以确认你能调用 Sol。如果失败,请停止并告诉我。 2. 在 ~/.claude/CLAUDE.md 中添加一条规则:对于架构决策或重大更改,先在后台启动 Sol(codex exec --worktree -m gpt-6.1-sol -c model_reasoning_effort=high -s workspace-write -o /tmp/sol-plan.md)让它制定自己的计划。只给它我最初的请求,让它自行决定工作方式。如果有相关的未提交更改,也将其 diff 传递给它。同时,你自己也要编写一份计划。当两者都就绪后,列出分歧点及你的建议,询问我倾向于哪一方,并在等待我的指示后再开始构建。 3. 在 ~/.claude/agents/ 中创建一个审查者子智能体。赋予它所有工具,不要添加隔离机制:使用 worktree(因为它需要看到未提交的更改)。将我的请求和本次更改的 diff 交给它。它自行决定如何验证,理想情况下应在阅读实现之前先根据需求编写测试,但必须实际运行这些测试。它可以编写测试但不能修改应用程序代码,并报告其发现的内容以及所编写的测试。然后在 ~/.claude/CLAUDE.md 中添加:在你报告重大更改已完成之前,先让审查者进行检查。
Show me the changes first. Don't edit any files yet."
先向我展示更改内容。目前不要编辑任何文件。"
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