Anthropic 员工如何使用 Claude Tag
英文标题：How Anthropic employees use Claude Tag
Claude Tag 可将 Claude 接入 Slack 等聊天工具，用户可在讨论串中 @Claude，让它理解上下文、完成任务并将结果回发至线程。Anthropic 员工已将其用于从 Slack 讨论生成营销文档、汇总分散的客户功能请求，以及加速营销内容的法律审查。案例显示，Claude Tag 可在约 26 分钟内汇总约 24 个相关客户账户，也能在约 50 分钟内将约 120 条原始发现整理为 23 个未解决问题和 14 个已解决问题。对开发者和企业团队而言，该工具的价值在于将跨频道检索、信息综合和初步审查等高耗时工作交给 Agent，同时保留人类对准确性、来源和最终决策的把关。
8.1 KB
428.7 KB
493.1 KB
622.6 KB
585.8 KB
403 B
17.7 KB
9.9 KB
15.1 KB
9.6 KB
41.4 KB
中文
Claude Tag 将 Claude 带入 Slack 等聊天工具：你可以像 @ 一位同事一样在讨论串中 @Claude，它会获取对话上下文、完成任务，并将答案或结果发回该讨论串。它还可以跟进对话，并利用可用上下文、其记忆以及已获得的常驻指令，决定何时参与聊天。过去数月中，Anthropic 的团队一直利用 Claude Tag 在共享频道中进行自助式数据分析、处理支持工单，或协助定位棘手 Bug 的根本原因。
我们基于 Anthropic 内部的工作经验，整理了十多个 Claude Tag 用例示例，其中包括具体提示词和设置说明。本文重点介绍 Anthropic 员工如何通过 Claude Tag 提高工作流和流程效率的三种方式，并提供他们使用的提示词，以便你借鉴或调整为最适合自己工作的版本。
将 Slack 讨论串转化为一份完善文档
在最近一次功能发布期间，一位销售代表希望获得一份面向客户和潜在客户、解释该功能工作方式的非技术材料。产品营销团队的 Hema Thanki 将随后产生的 Slack 讨论串转化为一份可供审核的文档，耗时 45 分钟。
该 Slack 讨论串包含超过 15 条消息，多人提出建议或额外需求；对于实际需要什么、现有技术材料是否足够，讨论中还存在一些分歧。Hema 没有试图澄清这些模糊之处，而是在讨论串中 @Claude：@Claude, go through this Slack thread and come up with a one pager that [the requester] is asking for.
要求 Claude 根据 Slack 讨论串生成营销材料。此图为说明用例而生成，不包含真实信息。
Claude 在约两分钟内生成了一份两页的草稿，涵盖该功能以通俗语言说明的作用、其商业价值、实施涉及的事项，以及附有更详细信息的附录。
接着，Hema 要求 Claude 核验其回答：@Claude, is everything in this doc factual and correct? Claude 将文档中的主张分为两类：一类已根据公开文档验证，另一类则属于它自身的表述框架；后者被标记为需要产品负责人确认。Hema 提供了两份包含相关信息的官方资源，Claude 随后重写了一个章节，使其措辞符合这些资源中已经批准的表述。
对 Claude 的后续指令同样在 Slack 中进行。此图为说明用例而生成，不包含真实信息。
此时，Hema 注意到被截断的讨论串使表述框架产生偏差，于是粘贴了更完整的上下文。她与 Claude 来回迭代，最终共形成四个版本。在首次提出请求约 45 分钟后，她将文档分享给该功能的产品负责人审核。Hema 不必花数小时研究和起草文档，而是将时间用于质疑准确性、提供来源，以及决定应纳入哪些信息；这些任务需要人类判断，也使这份客户材料更为扎实。
除根据 Slack 讨论串生成简报外，Hema 还在日常工作中广泛使用 Claude Tag。她创建了一个与 Claude 互动的私密 Slack 频道，在不同讨论串中提出请求，并像 @ 同事一样 @Claude。在该频道中，Claude 会阅读她粘贴或附上的任何内容，搜索 Slack 工作区和公开文档，并在后台工作，同时发布并持续更新进度清单。Claude 的访问权限经过刻意限定：它只会使用其被授予访问权限的频道和文档；当无法访问这些资源时，它也会告知她。
汇总并跟进散布在 Slack 各频道中的请求
当新功能发布时，销售代表通常需要追踪并向自己服务的、曾提出该功能需求的客户传达相关信息。这些请求通过 Slack 或产品反馈中心传达，并且可能分散在数月的历史记录中。产品战略与运营团队的 Steph Soderborg 在约 26 分钟内汇总了与一项即将发布功能相关的所有请求，并直接通知了每一位代表客户提出过该请求的销售代表。
首先，Steph 向 Claude 发送了搜索目标、关于何为匹配项的一句定义，以及一份粘贴的预期输出示例——此前一次发布时使用的七条条目列表：@Claude We are about to GA [a new feature]. Can you search Slack ... find me anyone who has asked for this functionality for their customer ... include their Slack handle and team, the account that asked for this, and link the ask from Slack.
Claude 在多个频道和更广泛的工作区内运行了约 20 种搜索变体。产品反馈中心阻止了它的直接访问，因此它改为通过 Slack 交叉引用呈现反馈中心的条目；同时，它还纳入另一名内部助手先前发布的一份初步列表，并对两份列表进行了去重。约 26 分钟后，Claude 返回了汇总列表，其中涉及约 24 个账户；每位请求者各占一行，包含其 Slack 用户名、团队、账户以及指向原始请求的链接。
要求 Claude 生成一份曾在 Slack 中发布某项特定功能请求的所有人员列表。此图为说明用例而生成，不包含真实信息。
随后，Steph 让 Claude 承担了一项更大规模的汇总工作——这项工作她自己没有足够的时间和精力完成：她希望了解企业客户在前一周报告过的每一个产品问题，包括哪些内容发生故障、哪些已经修复，以及哪些报告（如有）指向同一个底层问题。她要求 Claude 阅读所有涉及事故、升级处理、支持和产品反馈的 Slack 频道；约 50 分钟后，Claude 发布了一份按产品领域组织的报告，其中包括 23 个仍未解决的问题和 14 个已解决问题，这些内容由约 120 条原始发现压缩整理而来。每个问题都包含摘要及源讨论串链接。之后 Steph 要求 Claude 检查其工作，Claude 又发现了 15 个问题。
Steph 估计，梳理、分析并综合如此大量的信息，至少需要她全职工作一周，或者根本不会完成。借助 Claude Tag，她只需花几分钟明确自己的需求，Claude 则在后台完成工作。
Steph 也在一个私密频道中与 Claude 协作，在一开始就发送完整指令，其中包括应搜索哪些位置、何种情况算作匹配，以及通常还会给出输出格式示例。Claude 会搜索工作区、阅读被邀请进入的频道，并在工作期间发布进度更新。当反馈中心阻止访问时，Claude 会尝试通过可访问的文档和频道收集相关信息，甚至请求获得这些频道的访问权限。
加快法律文件审核
Anthropic 的法务团队会在每篇博客、落地页、电子邮件或其他任何材料公开发布前对其进行审核。在产品发布前的几天里，营销团队可能需要在紧迫期限内排队提交数十项不同资产供审核。这还不包括持续流入审核队列的其他营销材料，其范围从一段社交媒体文案，到 2,500 字博客草稿、包含十多个标签页的规划文档，以及在多个触点提供多个版本的电子邮件系列。法务团队的产品法律顾问 Molly Villagra 创建了一个专用 Slack 频道，由 Claude Tag 先检查每一项营销资产，将营销法务审核的周转时间从每项一天或更长时间，缩短至每项 30 分钟。
营销人员若要请求法务审核，只需在 Slack 频道中发布文档链接。Molly 没有工程技术背景，但已经为 Claude 设置了具体规则和指令。Claude 不仅能够发现法务问题，例如缺乏依据的营销声明；由于它可访问公司 Slack、内部知识索引和公开网络，它还可以帮助核查营销内容中的事实性表述。如果发现问题，Claude 会列出这些问题，并提供具体的处理说明，然后直接与请求者协作解决。对于仍需要法务签字确认的问题，Claude 会 @ 相应的产品法律顾问，使其能够快速审查被标记的表述。
营销人员通过在专用 Slack 频道中发布文档链接来请求法务审核。此图为说明用例而生成，不包含真实信息。
例如，在最近一次新闻通讯审核中，Claude 标记了三个关键事项，随后仅几分钟后，它在内部文档中找到所需信息，便在没有额外提示的情况下自行解决了其中一项。Molly 在营销法务审核频道中通过 @Claude，要求将这一行为设为默认规则：“ Your three bullets are good callouts, but they can all be verified by you. Will you try to verify these things in real time when you flag them in the future?” 在 Molly 的要求下，Claude Tag 将这一新指令添加到未来所有审核均须遵循的指令集合中，从而能够通过频道反馈实时改进。
这一反馈循环启发 Molly 建立了一项新惯例：她指示 Claude 在每周五审查当周法律顾问的反馈，并提出对共享指令的更新建议，供她批准。
我们在上文分享的每一种工作流，都为 Anthropic 员工节省了数小时或数天的工作时间，并促成了一些此前根本不会发生的项目。你的团队最先会将哪些工作流或项目交给 Claude？
目前处于公开测试阶段的 Claude Tag 已在 Anthropic 第一方服务的 Team 和 Enterprise 套餐中提供。你可以在 claude.ai/admin-settings/claude-tag 为工作区进行设置，或在 claude.com/docs/claude-tag 了解更多信息。
本文中的周转时间反映个别员工处理特定任务的经验；结果会因任务、接入的工具以及 Claude Tag 的设置方式而异。
所有图片均为说明用例而生成，不包含真实姓名或信息。
English / 原文
Claude Tag brings Claude into chat tools like Slack, where you can tag @Claude in a thread the way you would a colleague and it picks up the context of the conversation, completes the task, and posts the answer or results back in the thread. It can also follow conversations and draw on available context, its memory, and standing instructions it’s been given to decide when to participate in the chat. Over the past several months, teams at Anthropic have been using Claude Tag to self-serve data analysis in shared channels, work through support tickets, or help find the root cause of tricky bugs.
We’ve assembled more than a dozen use case examples for Claude Tag inspired by our work at Anthropic, along with specific prompts and setup instructions. In this post, we highlight three ways Anthropic employees are making their workflows and processes more efficient with Claude Tag, with the prompts they used, so you can borrow or adapt the ones that best fit your work.
Turning a Slack thread into a polished document
During a recent feature launch, a sales rep asked for non-technical collateral that explains how the feature works to customers and prospects; Hema Thanki, on the product marketing team, turned the Slack thread that followed into a review-ready document in 45 minutes.
That Slack thread ran to more than 15 messages, with multiple people chiming in with suggestions or additional asks, and a touch of tension around what was actually needed and whether the existing technical material was enough. Rather than attempting to clarify ambiguity, Hema tagged Claude in the thread: @Claude, go through this Slack thread and come up with a one pager that [the requester] is asking for.
Asking Claude to generate marketing collateral based on a Slack thread. This image has been generated to illustrate a use case and does not contain real information.
Claude generated a two-page draft in about two minutes, covering what the feature does in plain terms, the business case for it, what implementation involves, and an appendix with more detailed information.
Next, Hema asked Claude to verify its responses: "@Claude, is everything in this doc factual and correct?" Claude sorted the document's claims into ones verified against public documentation and those that were its own framing, which it flagged for product-lead sign-off. Hema supplied two official resources with relevant information, and Claude rewrote one section to match the approved wording in those resources.
Follow-up instructions for Claude also happen in Slack. This image has been generated to illustrate a use case and does not contain real information.
At that point, Hema noticed the truncated thread had skewed the framing, so she pasted in the fuller context. She went back and forth with Claude for a total of four versions. About 45 minutes after the first ask, she shared the document with the feature's product lead for review. Rather than spend hours researching and drafting the document, Hema’s time went to challenging accuracy, supplying sources, and deciding what information to include, all tasks that required human judgment and made the customer asset even stronger.
Beyond generating briefs from Slack threads, Hema also uses Claude Tag across her day to day work. She keeps a private Slack channel with Claude where she makes requests in separate threads, @-mentioning Claude the way she'd tag a colleague. In that channel, Claude reads whatever she pastes or attaches, searches the Slack workspace and public documentation, and works in the background, posting a progress checklist it updates as it goes. Claude’s access is deliberately scoped: it only works from the channels and documents it has been granted access to, and will let her know when it does not have the access to these resources.
Consolidating and following up on requests scattered throughout Slack channels
When a new feature launches, sales reps typically keep track and communicate it to customers they support who have requested that feature. Those requests are communicated via Slack or in a product feedback hub, and can be scattered across months’ worth of history. Steph Soderborg, on the product strategy and operations team, was able to consolidate all asks related to an upcoming feature and directly notify each rep who had asked for it on behalf of a customer, in about 26 minutes.
To start, Steph messaged Claude with the search targets, a one-sentence definition of a match, and a pasted example of the output she wanted, a seven-entry list from a previous launch: "@Claude We are about to GA [a new feature]. Can you search Slack ... find me anyone who has asked for this functionality for their customer ... include their Slack handle and team, the account that asked for this, and link the ask from Slack."
Claude ran about 20 search variants across several channels and the wider workspace. The product-feedback hub blocked its direct access, so it surfaced hub items through Slack cross-references instead, and it folded in a first-pass list another internal assistant had posted, deduplicating the two. The consolidated list came back in about 26 minutes and included roughly 24 accounts, with one line per requester containing their Slack handle, team, account, and a link to the original ask.
Asking Claude to generate a list of everyone who has posted asked for a specific feature in Slack. This image has been generated to illustrate a use case and does not contain real information.
Steph then put Claude on a bigger consolidation job that she wouldn’t have had the bandwidth to do on her own: she wanted a picture of every product problem enterprise customers had reported in the previous week, including what was broken, what was already fixed, and which, if any, reports pointed at the same underlying issue. She told Claude to read all Slack channels covering incident, escalation, support, and product-feedback, and roughly 50 minutes later Claude posted a write-up, organized by product area, that included 23 issues that were still open and 14 resolved ones, condensed from about 120 raw findings. Each issue included a summary and a link to the source thread. Steph then asked Claude to check its work, and it surfaced 15 more issues.
Steph estimates that combing through, analyzing, and synthesizing this much information would have taken her at least a week of full-time work, or would never have gotten done. Instead, with Claude Tag, she took a few minutes to shape up her ask, and Claude worked in the background.
Steph also works with Claude in a private channel, sending full instructions up front that include where to search, what counts as a match, and usually an example of the output format. Claude searches the workspace, reads the channels it has been invited to, and posts progress updates as it works. When the feedback hub blocks access, Claude attempts to gather related or relevant information via accessible docs and channels, or even asks for access to these channels.
Expediting legal document reviews
Anthropic’s legal team reviews each blog, landing page, email, or any other collateral before it’s publicly released. In the days leading up to a product launch, the marketing team can queue up dozens of different assets for review, on a tight deadline. That’s on top of all other marketing collateral flowing through the review queue, ranging from one-paragraph social copy to 2,500-word blog drafts, planning documents with a dozen-plus tabs, and email series with multiple variants across multiple touchpoints. Molly Villagra, a product counsel on the legal team, created a dedicated Slack channel where Claude Tag examines every marketing asset first, compressing marketing legal review turnaround time from a day (or longer) to 30 minutes per asset.
To request legal review, marketers post a document link in the Slack channel, where Molly, who has no engineering background, has set up specific rules and instructions for Claude. Not only can Claude spot issues for legal (like unsubstantiated marketing claims), but it can also help check factual statements in the marketing content because it has access to the company Slack, an internal knowledge index, and the public web. If there are flags, Claude lists those with specific instructions on how to address them and works directly with the requester to do so. For remaining issues that need legal sign-off, Claude tags the appropriate product counsel, who can quickly review the flagged statements.
Marketers request legal review by posting a link to the document in a dedicated Slack channel. This image has been generated to illustrate a use case and does not contain real information.
In a recent newsletter review, for example, Claude flagged three key items, then just minutes later, unprompted, resolved one of them after finding the information it needed in internal documents. Molly asked it to make this the default by tagging @Claude in the marketing legal review channel: “ Your three bullets are good callouts, but they can all be verified by you. Will you try to verify these things in real time when you flag them in the future?” At Molly’s request, Claude Tag added this new instruction to its set of instructions to follow in all future reviews, allowing it to improve with channel feedback in real time.
This feedback loop inspired Molly to create a new routine, instructing Claude to review the week’s counsel feedback each Friday and propose an update to the shared instructions for her approval.
Each of the workflows we’ve shared above is saving Anthropic employees hours or days of work, and enables projects that simply wouldn’t have happened before. What workflows or projects would your team hand over to Claude first?
Claude Tag, currently in public beta, is available on Team and Enterprise plans, on Anthropic’s first-party service. Set it up for your workspace at claude.ai/admin-settings/claude-tag or learn more at claude.com/docs/claude-tag.
Turnaround times in this post reflect individual employees' experiences with specific tasks; results vary with the task, the tools connected, and how Claude Tag is set up.
All images have been generated to illustrate use cases and do not contain real names or information.