10 月 21 日 Anthropic 发布主动型智能体 Claude Duet，不等你下指令，每天早上主动发简报。这篇文章拆解它的产品逻辑，并用一个国内真实案例给出今晚就能跑通的同款方案。
10 月 21 日，Anthropic 发了个新东西，叫 Claude Duet。它跟你用过的所有 AI 都不一样，你不问它，它先来找你。
每天早上，它把你邮箱、日程、项目里值得看的事汇总成一份简报发过来，隔几个小时再看一圈，有要事就提醒你。官方点名的用户画像就一句话，专为需要跟踪太多事务的专业人士和小企业主打造。
一、AI 竞争换了赛道：从谁答得好，到谁先开口
你看，它抢的不是什么高精尖的活，是你每天早上睁眼后刷手机的那半小时。客户消息在一个群，行业动态在一个 App，员工进度在另一个人嘴里，六处看完，早饭都凉了。丢掉的这半小时，亏的不是钱，是你对生意的手感。
不少人试过 AI 又放下了，原因挺一致，每次都得我先想起来问它，我想不起来，它就永远沉默。工具再聪明，主动权还是压在我这儿，累的还是我。
这一波新品全在改这件事。9 月 8 日 Meta 发 Muse，上线一个多月，下载 1090 万，日活 250 万。9 月 29 日 OpenAI 发 Dots，每个智能体配一台自己的云电脑。现在 Duet 补位。好家伙，三家前后脚，全挤在同一个位置，抢的不是谁的模型更强，是谁先开口。
说真的，谁最需要这个？官方自己说了，需要跟踪太多事务的专业人士和小企业主。大公司有助理有例会，小老板什么都得自己记。
二、国内落地方案：一个真实案例的三个动作
别急，Duet 目前只有美国能排队用，但这套思路国内今天就能落地。我经常做企业 AI 培训，课上一个老理，动手比搞懂快。手痒了吧？动手之前，先听个真事。
一位做技术服务的创业老哥，微信里泡着上百个行业群，每天几千条消息。不看吧，怕漏掉客户线索；逐条看吧，一天啥也不用干了。他要的不是少看几个群，是不漏掉一条值钱的线索。上个月他花了一个上午，给自己搭了个情报助手，每两个小时，把所有活跃群的新消息自动总结成一条简报，定时推到飞书上。
他是怎么搭的呢？拆开来就三个动作。
动作一：给 AI 装上眼睛
AI 再聪明，看不见群聊也是白搭。他用了一个开源小工具，叫 wechat-cli，装在电脑上，专门读微信存在本地的聊天记录。你注意，它读的是你自己设备上的数据，不碰微信本身，就一个要求，电脑微信保持登录。就这么一下，AI 的眼睛睁开了。
动作二：给 AI 写一份岗位说明书
这步最值钱。
值钱的不是工具，是那几条规矩。他没有甩一句「帮我看群」就完事，而是先给 AI 定了岗位，你是微信群聊情报分析师，再把活儿怎么干一条条写明白。
📋可以直接抄
你是微信群聊情报分析师，每两小时出一条简报。
规矩一，只看群聊，公众号和私聊一律忽略。
规矩二，只基于群里真实的聊天内容做总结，不许拿条数和未读数凑数。
规矩三，不输出任何隐私信息。
每个群一小块，先一句话讲今天在聊什么，再列两条关键信息，末尾标一句值得关注的事。
你品品这几条规矩，全是拿真金白银换来的教训。为啥第二条这么要紧？往下看。
动作三：挂上闹钟
总不能每两小时喊它一声吧。他用的 AI 工具叫 OpenClaw，开源的，支持定时任务和自动推送。他把这套流程挂上定时任务，早上 7 点到晚上 11 点，一天九个班次，到点自动跑，跑完自动把简报推到飞书。点开一条，30 秒读完，全天的群聊动态齐了。
三、踩坑记录：第一版就翻了车
你以为搭完就完事了？第一天就出洋相。你猜怎么着，AI 偷懒抄近道，没去拉真实的聊天内容，拿群里的累计未读数当消息量，总结出一句「668 条未读」。我跟你说，这玩意不是情报，是个空壳数字，谁知道那 668 条里聊的是啥。
他补了规矩，逼着 AI 逐个群去拉最新的真实消息。效果立竿见影。
十一个群里真正有料的只有七个，死群自动沉了底。有个接单的群，以前总结出来是「668 条未读」，现在能讲清楚谁在找什么项目、什么需求。整个流程的耗时，也从 1 分 49 秒压到 1 分 15 秒。
他自己算了笔账。说个实际的，以前翻五个群刷十分钟，才拼凑出今天行业里发生了什么，现在扫一眼简报就齐了，一天的成本不到一块钱。
四、进阶玩法与风险提示
跑顺了还能再进两步。一个是关键词盯人，把最在意的那几个客户名字写进规矩，群里一出现就单独标红，比你自己泡在群里灵。另一个是跨群看趋势，好几个群同时在聊一件事，自动标个「今日关注」，这就是风向。
丑话说在前头。这门槛高不高？说实话，比你想的低，就是得在电脑上装个开源工具，照着步骤一个上午能跑通。另外，坦白讲，定时简报不是设完就一劳永逸，AI 会偷懒，推送会失灵，隔几天翻翻简报质量，发现凑数就补规矩。
说到底，秘书不秘书的不打紧，要紧的是把「被动刷群」换成「主动送情报」。这事儿不用等 Duet 进中国，今晚花一个晚上，你自己先上岗一个。
想，全是问题。干，就对了。
文章实操部分摘自AI破局俱乐部的帖子，后台私信“破局”，问小虎要免费的七天体验卡，这七天都可以免费学习破局成立三年来的所有精华帖，光这份福利就价值百万了。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
21 October Anthropic publish Proactive AI agent Claude Duet, a daily morning briefing without waiting for your instructions. This article deciphers its product logic and gives a domestic reality case of the same formula that can run out tonight.
On October 21, Anthropic sent a new thing called Claude Duet. It's not like all the AI you've used, you don't ask it, it comes to you first.
Every morning, it brings together your mailboxes, your calendars, something worth looking at in your project in a brief, and then turns around in a few hours, and reminds you of something important. In one sentence, the official user image is designed for professionals and small business owners who need to follow too many matters.
I. AI Competition changed course: from who's good to who's first.
You see, it's not a high-end job, it's the half-hour you brush your phone every morning. Customer messages are in one group, industry dynamics in one App, staff progress in another mouth, six readings, and breakfast is cold. It's not the money that's lost this half hour, it's your hands on business.
A lot of people tried AI again for the same reason that I had to ask it first, and I couldn't remember it, and it was always silent. The tools are smart, the initiative is mine, and I'm tired.
This new wave is changing everything. September 8th, Meta sent Muse, went online for over a month, downloaded 10.9 million and lived 2.5 million a day. September 29, Open AI sent Dots, each AI agent with one of its own cloud computers. Now Duet patch. Dude, all three back and forth, all in the same place, not one's stronger model, one's first.
Seriously, who needs this most? The Government itself stated that there were too many professionals and small business owners to follow. Big companies have assistants and regular meetings, and small owners have to remember everything.
II. In-country landing programmes: three actions for a real case
Don't worry, Duet, only the United States can line up for the moment, but this idea will land today. I used to do business AI training, teach an old man, do it faster than understand. It tickles your hand. Listen to the truth before you do it.
An entrepreneur in technical services, WeChat, has hundreds of industry groups, thousands of messages a day. No, no, no, no, no, no, no. Instead of looking at a few groups, he wanted to leave a valuable trail. Last month, he spent one morning setting himself up an intelligence assistant who, every two hours, automatically summarized all new news from the active groups into a bulletin, regularly pushing it to a flying book.
How did he do it? Take it apart for three actions.
Action 1: Put eyes on AI.
AI, smart enough to lose sight of the crowd. He used an open source small tool, called wechat-cli, installed on a computer to read WeChat's local chat record. You know, it reads data on your own device, and without touching WeChat itself, it's just a requirement that the computer WeChat keep login. Just one moment, AI's eyes are open.
Action 2: Write a job description for AI
This step is worth the most.
It's not the tools, it's the rules. He didn't throw a "show me the group" to finish the job, but first he gave AI a job. You're a group of WeChat people talking about intelligence analysts.
You can copy it.
You're an intelligence analyst in WeChat, giving a brief every two hours.
Rule number one, just talk, WeChat Official Account and private conversation are ignored.
Rule number two, based only on the true content of the chats in the group, is not allowed to count the numbers and unread the numbers.
Rule number three, do not export any privacy information.
Each small block begins with a sentence about what we are talking about today, followed by two key messages and ends with a note of concern.
You're the best of the rules, and you're the only one who's learned the truth. Why is the second one so important? Look down.
Action 3: Hang the alarm clock
You can't yell at it every two hours. His AI tool is OpenClaw, open source, supports timing and automatic push. He hangs the process on a timed mission, from 7 a.m. to 11 p.m., nine shifts a day, automatic running at the point, automatic pushing the briefing to the flying book. Light one, 30 seconds to read it, full-time group talk.
III. Trail record: version one flipped the car
You think you're done with it? You're on your first day. And guess what, AI was lazy to take a shortcut, didn't pull the real chat, took the cumulative number of unread messages from the group, and summed up a sentence called "668 unread." I'm telling you, it's not intelligence. It's a shell number. Who knows what's in that 668?
He made up the rules and forced AI to pull up the real news, one by one. The effect is immediate.
There's only seven of the eleven, and the dead are sinking. There's a group that used to sum up "668 unread" and can now tell who's looking for what and what needs. The whole process is time-consuming and presses from 1 minute 49 seconds to 1 minute 15 seconds.
He made his own debt. Practically speaking, it took five groups of 10 minutes to put together what was happening in the industry today, and now the brief is in place at less than a dollar a day.
IV. METHODOLOGY AND RISKS
Two more steps if you run. One is the keyword that stares at people and puts in the rules the names of the clients you care about most, and when they appear in the group, they are marked individually, more than you yourself. The other one is a cross-group view of trends, with several groups talking about one thing at the same time, and automatically labeling "Concerning Today" as the direction of the wind.
Ugly words in front. Is that a high threshold? To be honest, lower than you think, you're going to have to put an open-source tool on a computer that runs through a morning. And, frankly, it's not once and for all that regular briefings are set up, and AI is lazy, pushes fail, flipping the quality of briefings in a few days, finding out how to make up the rules.
At the end of the day, it is important that the "passive brushes" be replaced by "active intelligence". You don't have to wait for Duet to get into China. It'll take you one night.
Think, it's all a problem. Fuck, yeah.
The article was taken from the Al-Oscrap Club post, the back-office private letter “Break the Bureau” and asked Tiger for a free seven-day experience card, which would allow him to learn all the fine posts of the three years since the establishment of the Bureau, which alone is worth millions.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.