Anthropic发布Claude Code Projects完整指南及Opu
Anthropic just shared its 𝗳𝘂𝗹𝗹 𝗴𝘂𝗶𝗱𝗲 to Claude Code Projects, and every Pro and…
提供了可直接复用的复杂Prompt模板，展示了如何将官方文档转化为具体的自动化工作流，结构清晰且具备实操参考价值。
Anthropic just shared its 𝗳𝘂𝗹𝗹 𝗴𝘂𝗶𝗱𝗲 to Claude Code Projects, and every Pro and Max user on the waitlist got in today. You can have Opus 5.5 follow it to go through your recent sessions and turn 𝘁𝗵𝗲 𝗯𝗮𝗰𝗸𝗴𝗿𝗼𝘂𝗻𝗱 𝘆𝗼𝘂 𝗿𝗲-𝗲𝘅𝗽𝗹𝗮𝗶𝗻 every time you start a new session into a set of project instructions.
Anthropic 刚刚分享了其 𝗳𝘂𝗹𝗹 𝗴𝘂𝗶𝗱𝗲（完整指南）关于 Claude Code Projects，等待名单上的所有 Pro 和 Max 用户今天都已获得访问权限。你可以让 Opus 5.5 跟随该指南，浏览你最近的会话，并将每次启动新会话时都需要重新解释的 𝘁𝗵𝗲 𝗯𝗮𝗰𝗸𝗴𝗿𝗼𝘂𝗻𝗱 𝘆𝗼𝘂 𝗿𝗲-𝗲𝘅𝗽𝗹𝗮𝗶𝗻（你需要重新解释的背景信息）转化为一组项目指令。
In a new project, every thread runs on Opus at high effort by default, which uses up your plan fastest. When you hit your limit, a thread doesn't end. It waits for the limit to reset and keeps going on its own. Without a word from you, 𝗶𝘁 𝘀𝘁𝗮𝗿𝘁𝘀 𝘂𝘀𝗶𝗻𝗴 𝘆𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝘂𝘀𝗮𝗴𝗲 𝘄𝗶𝗻𝗱𝗼𝘄. So before you send a batch of work, check Thread model and Thread effort in the project settings, and give work that doesn't need Opus a smaller model or lower effort.
在新项目中，默认情况下每个线程都以高努力级别在 Opus 上运行，这会最快消耗你的计划配额。当你达到限制时，线程不会结束。它会等待限制重置并自行继续。无需你发出任何指令，𝗶𝘁 𝘀𝘁𝗮𝗿𝘁𝘀 𝘂𝘀𝗶𝗻𝗴 𝘆𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝘂𝘀𝗮𝗴𝗲 𝘄𝗶𝗻𝗱𝗼𝘄（它将开始使用你的下一个用量窗口）。因此，在发送一批工作之前，请检查项目设置中的 Thread model（线程模型）和 Thread effort（线程努力级别），并为不需要 Opus 的工作分配较小的模型或较低的努力级别。
Send this prompt to Opus 5.5 👇
将此提示发送给 Opus 5.5 👇
"Read this doc: https://code.claude.com/docs/en/claude-projects
"阅读此文档：https://code.claude.com/docs/en/claude-projects
Then check my last 20 sessions (the .jsonl files in this repo's folder under ~/.claude/projects/, only what I typed myself) and my recent commits against it, and find three kinds of things: 1. Work that keeps producing new tasks: bugs fixed again and again in the same code, or migrations and features that took several sessions to finish. Using the doc's 'When to use a project' section, pick the one most worth turning into a project and explain why. Based on how hard it is, tell me how to set Thread model and Thread effort. Then pick one small, real task and write it as a first message I can send straight into the project conversation. 2. Background I re-explain every time I start a new session: the target branch, how to run tests, how to name PRs, what to ask me about first. List how often I said each one, with my exact words. Following the five points in the doc's 'Write project instructions' section, write instructions for that project that I can paste straight into Project instructions. Mark anything that only applies to this repo, like build commands, for the repo's CLAUDE.md. 3. Things I keep correcting you on: anything I've said two or more times that isn't already in the instructions. Quote my exact words and turn each into one line I can send straight into the project conversation: 'Remember: …'
然后检查我最近的 20 个会话（位于 ~/.claude/projects/ 下该仓库文件夹中的 .jsonl 文件，仅包含我自己输入的内容）以及针对该文档的最新提交，并找出三类内容： 1. 持续产生新任务的工作：反复修复同一代码中的 bug，或需要多个会话才能完成的迁移和功能。根据文档中“何时使用项目”部分，选择最值得转化为项目的任务并解释原因。根据其难度，告诉我如何设置 Thread model（线程模型）和 Thread effort（线程努力级别）。然后选择一个小型的真实任务，将其编写为第一条消息，我可以直接将其发送到项目对话中。 2. 每次启动新会话时我都会重新解释的背景：目标分支、如何运行测试、PR 的命名规范、首先应该向我询问什么。列出每项内容的出现频率，并附上我的原话。遵循文档“编写项目指令”部分的五个要点，为该项目的 Project instructions（项目指令）编写可直接粘贴的指令。标记仅适用于此仓库的内容（如构建命令），以便放入仓库的 CLAUDE.md 中。 3. 我不断纠正你的事项：任何我已说过两次或以上且尚未包含在指令中的内容。引用我的原话，并将每条内容转化为一行我可以直接发送到项目对话中的指令：'Remember: …'（记住：……）
For every item, quote my words or the commit, and say what I'd no longer have to repeat or fix once it's in place.
对于每一项，引用我的话或提交记录，并说明一旦实施后，我不再需要重复或修正的内容。
Show me the list and the project instructions first. Don't change anything yet."
先向我展示列表和项目指令。暂时不要进行任何更改。"
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