跳到主内容
@wquguru
精选85VoxAI 产品与模型

Claude Code 使用技巧:利用 Haiku 5.5 子代理优化

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝘁𝗶𝗽: Haiku 5.5 came out yesterday. Keep Opus 5.5 as your main model…

原文
发到 X
推荐理由

提供了具体的模型分工策略与可直接复用的配置 Prompt,结构清晰且有实测数据支撑,对开发者有实操参考价值。

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝘁𝗶𝗽: Haiku 5.5 came out yesterday. Keep Opus 5.5 as your main model and hand work like reading code and looking up docs to Haiku 5.5 subagents.

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 技巧:Haiku 5.5 昨天发布了。将 Opus 5.5 保留为主模型,并将阅读代码和查阅文档等工作交给 Haiku 5.5 子代理处理。

For the launch, Anthropic ran an egg-drop test on Claude Managed Agents. Opus 5.5 with 10 Haiku 5.5 subagents hit the 32-meter goal in 𝟱𝟴 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗳𝗼𝗿 $𝟬.𝟭𝟰. Opus alone took 3:37 and $0.47.

为了庆祝发布,Anthropic 对 Claude Managed Agents 进行了鸡蛋掉落测试。使用 10 个 Haiku 5.5 子代理的 Opus 5.5 在 𝟱𝟴 秒内、花费 $𝟬.𝟭𝟰 完成了 32 米的目标。而单独使用 Opus 则耗时 3:37,花费 $0.47。

One default behavior needs changing too. Claude Code's built-in Explore subagent follows your main model, so when Opus 5.5 sends Explore to search your code, 𝗘𝘅𝗽𝗹𝗼𝗿𝗲 𝗿𝘂𝗻𝘀 𝗼𝗻 𝗢𝗽𝘂𝘀 𝗮𝘀 𝘄𝗲𝗹𝗹. Create your own subagent named Explore with model: haiku, and it replaces the built-in one.

还有一个默认行为需要更改。Claude Code 内置的 Explore 子代理会跟随你的主模型运行,因此当 Opus 5.5 发送指令让 Explore 搜索代码时,𝗘𝘅𝗽𝗹𝗼𝗿𝗲 也会 𝗿𝘂𝗻𝘀 𝗼𝗻 𝗢𝗽𝘂𝘀(在 Opus 上运行)。创建一个名为 Explore 且 model: haiku 的自定义子代理,它将替换内置的那个。

The tree is the full setup. Opus 5.5 on high runs the main session. Three subagents on medium split the rest: Explore and the researcher move to Haiku 5.5, one reading code and one looking up docs, while the worker that edits code and runs tests stays on Opus 5.5. Hand the tree and this prompt to Claude Code to set it up 👇

这是完整的配置结构。高负载下的 Opus 5.5 运行主会话。三个中等负载的子代理分担其余工作:Explore 和研究员迁移到 Haiku 5.5,一个负责阅读代码,一个负责查阅文档;而负责编辑代码和运行测试的工作者则留在 Opus 5.5 上。将此结构和提示词交给 Claude Code 以进行设置 👇

"Set up my Claude Code to match this tree: 1. Check claude --version. If it's below 2.1.293, stop and tell me to run claude update, since Haiku 5.5 needs it. If I'm on Bedrock or Vertex, tell me, because haiku has to be a full model ID there. 2. For the three subagents, reuse fitting ones from ~/.claude/agents and .claude/agents, and propose new ones in ~/.claude/agents only for missing roles. Explore (replacing the built-in one) and the researcher get model: haiku and effort: medium. Explore gets only Read, Grep and Glob; the researcher also gets WebFetch and WebSearch. The worker gets model: opus and effort: medium and handles edits and tests. Give each subagent a description that says what work to hand it, and write Explore's to match the built-in one's purpose. If any already sets a different model, or the project has one with the same name, list it and ask me before changing it. 3. In ~/.claude/settings.json, set the main session's default model to Opus 5.5 and its default effort to high. If I've saved a different effort level for Opus 5.5, tell me. 4. Check whether CLAUDE_CODE_EFFORT_LEVEL is set, since it overrides every subagent's effort. Report it; don't change it. Show me the changes first. Don't edit any files yet."

"按照此树状结构设置我的 Claude Code: 1. 检查 claude --version。如果版本低于 2.1.293,请停止并告诉我运行 claude update,因为 Haiku 5.5 需要该版本。如果我使用的是 Bedrock 或 Vertex,请告诉我,因为在那里 haiku 必须是完整的模型 ID。 2. 对于这三个子代理,从 ~/.claude/agents 和 .claude/agents 中复用合适的现有代理,仅针对缺失的角色在 ~/.claude/agents 中提议新建代理。Explore(替换内置的那个)和研究员使用 model: haiku 和 effort: medium。Explore 仅拥有 Read、Grep 和 Glob 权限;研究员额外拥有 WebFetch 和 WebSearch 权限。工作者使用 model: opus 和 effort: medium,负责编辑和测试。为每个子代理编写描述,说明应分配给它们的工作内容,并为 Explore 编写的描述需匹配内置代理的目的。如果任何现有代理设置了不同的模型,或者项目中存在同名代理,请先列出并询问我,然后再进行修改。 3. 在 ~/.claude/settings.json 中,将主会话的默认模型设置为 Opus 5.5,默认努力程度(effort)设置为 high。如果我之前为 Opus 5.5 保存了不同的努力程度,请告诉我。 4. 检查是否设置了 CLAUDE_CODE_EFFORT_LEVEL,因为它会覆盖所有子代理的努力程度。报告其状态;不要修改它。 先向我展示更改内容。暂时不要编辑任何文件。"

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

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件