九月十七日

Anthropic Claude Code builder Boris Cherny says Claude chat and Cowork are beginning a gradual merger toward one experience that carries context across tasks. Claude Docs, Claude Slides, and Claude Design are also moving into conversations, with editable presentations that can be exported as PowerPoint or PDF.

Anthropic Claude Code 开发者 Boris Cherny 表示，Claude chat 与 Cowork 正逐步合并，目标是在不同任务之间延续上下文。Claude Docs、Claude Slides 和 Claude Design 也将进入对话，用户可以编辑生成的演示文稿，并导出为 PowerPoint 或 PDF。

[Source 1](https://x.com/bcherny/status/2100260544087535639) · [Source 2](https://x.com/bcherny/status/2100259951398789487)

Anthropic Claude Code and Cowork builder Cat Wu explains that the merged experience lets Claude choose between a quick answer and deeper agentic work based on the prompt, including the appropriate output format. Users can still stop or redirect it and adjust its effort or approach; the rollout will take several weeks.

Anthropic Claude Code 与 Cowork 开发者 Cat Wu 解释，合并后的体验让 Claude 根据 prompt 判断是快速回答还是开展更深入的 agent 工作，并选择合适的输出形式。用户仍可随时停止、调整方向或控制投入程度与执行方式；这次更新会在数周内逐步推出。

[Source 1](https://x.com/_catwu/status/2100260655312089562)

Vercel CEO Guillermo Rauch reports that the safety reviewer in fx currently checks every command in auto mode using GPT Luna. He says TypeSafe AI’s Jev is up to 18 times faster at p95 and more accurate in their results, with AI Gateway availability planned and a possible switch in the default reviewer.

Vercel CEO Guillermo Rauch 表示，fx 的 auto 模式目前使用 GPT Luna 审查每一条命令。他称，团队测得 TypeSafe AI 的 Jev 在 p95 延迟上最高快 18 倍，准确率也更高；Jev 计划接入 AI Gateway，并可能成为默认审查模型。

[Source 1](https://x.com/rauchg/status/2100307962262872105) · [Source 2](https://t.co/y5tFnlFN2l)

Anthropic Claude Code builder Thariq has revised his view that bash alone is sufficient for reliable tool calling: give Claude tools matching the operation, such as a database API for storage, instead of adding a filesystem abstraction. He still favors bash and sandboxes for generating and executing code. Claude Managed Agents makes the sandbox optional and independently startable, and he reports successfully porting an older bash-tool project to it.

Anthropic Claude Code 开发者 Thariq 修正了“可靠工具调用只需要 bash”的看法：工具应直接匹配实际操作，例如存储数据时提供数据库 API，而不是额外套一层文件系统抽象。代码生成与执行依然适合使用 bash 和 sandbox。Claude Managed Agents 的 sandbox 可选，也能独立启动；他表示，已成功将一个旧的 bash 工具调用项目迁移过去。

[Source 1](https://x.com/trq212/status/2100315538472009897) · [Source 2](https://x.com/trq212/status/2100315537251463523) · [Source 3](https://x.com/trq212/status/2100315535758217422) · [Source 4](https://t.co/xTptTmEmCE)

Box CEO Aaron Levie sees fast, inexpensive, capable models as useful gates throughout enterprise agent workflows. He points to data classification, routing, domain-specific decisions, and safety or security judgments as concrete applications; the supplied post does not name the model behind his reaction.

Box CEO Aaron Levie 认为，速度快、成本低且能力足够的模型适合承担企业 agent 工作流中的各类判断环节。他列出的具体用途包括数据分类、流程路由、领域决策以及安全判断；所提供的帖子没有明确点出他评论的模型名称。

[Source 1](https://x.com/levie/status/2100448648672993540)

Meta Senior Director of AI Madhu Guru argues that safety and security are features of an AI product and model. His point is that they belong within the product itself, rather than being treated as externally imposed guardrails.

Meta AI 高级总监 Madhu Guru 认为，安全与防护本身就是 AI 产品和模型的功能。他强调，这些能力应当属于产品本身，而不是被当作外部强加的限制。

[Source 1](https://x.com/realmadhuguru/status/2100312717739667963) · [Source 2](https://t.co/jtN3rzdZV8) · [Source 3](https://t.co/J8dM9vHsoN)

AI tutorial and interview creator Peter Yang says stronger models have not made skills redundant in his podcast workflow, because skills preserve his editing and browser-use instructions. He uses eight skills: /podcast-prep researches guests and drafts interview guides, /podcast-edit helps select intro quotes and cuts, and /podcast-production coordinates five skills to produce six assets while he retains editorial judgment.

AI 教程与访谈创作者 Peter Yang 表示，模型变强并没有让播客工作流中的 skills 失去价值，因为它们能保留他特定的剪辑与浏览器操作要求。他使用八个 skills：/podcast-prep 研究嘉宾并准备采访提纲，/podcast-edit 协助选择开场引语和删减片段，/podcast-production 则编排五个 skills 生成六种素材，同时由他保留编辑判断。

[Source 1](https://x.com/petergyang/status/2100328939034128856) · [Source 2](https://t.co/PYNbe5tlFw)

FPV Ventures partner Nikunj Kothari describes a Nous Research “Home” agent that he and his wife use as a shared bot. He values granular control over selected emails, converting inline attachments into structured data, and logged-in browser sessions, and says he has not reproduced that control with Grok Bot and similar tools.

FPV Ventures 合伙人 Nikunj Kothari 分享了他与妻子共同使用的 Nous Research “Home”agent。他看重的是细粒度控制：只读取指定邮件、将内嵌附件转换为结构化数据，以及使用已登录的浏览器会话；他表示，尚未用 Grok Bot 等工具复现同等程度的控制。

[Source 1](https://x.com/nikunj/status/2100212813625196917)

Builder Zara Zhang says she is using Claude less because its speaking style feels more concerned with displaying sophistication than communicating clearly. Her criticism targets the way it expresses answers, rather than a specific task failure.

开发者 Zara Zhang 表示，自己最近减少了 Claude 的使用，因为它的表达方式更像是在展示复杂与高深，而不是把意思说清楚。她批评的是回答的表达方式，没有指出某次具体任务失败。

[Source 1](https://x.com/zarazhangrui/status/2100278750776824115)

#### Claude Cowork and chat are now one Claude

Claude Blog: “Claude Cowork and chat are now one Claude” announces a merged experience that can answer quick questions or continue larger tasks after the laptop is closed. The September 16 announcement starts with Pro and Max across web, desktop, and mobile over the coming weeks; Team and Free will follow. Enterprise administrators will receive at least 30 days’ notice before changes. The rationale is direct: “So we stopped making you choose.”

Claude Blog 的《Claude Cowork and chat are now one Claude》宣布合并后的体验：既能回答简短问题，也能在用户合上笔记本后继续处理较大的任务。这项 9 月 16 日公布的更新将在未来数周先覆盖 Pro 和 Max 的网页端、桌面端与移动端，Team 和 Free 随后跟进。Enterprise 管理员会在变更前至少 30 天收到通知。文章直接说明了这一调整的出发点：“所以，我们不再让你做选择。”

Claude Docs and Claude Slides are new, while Claude Design now works inside conversations; all three are in beta on paid plans, with Enterprise admins controlling activation. Users can edit directly, comment on slides, present from Claude, export PowerPoint or PDF, and share work through one link. Existing Cowork chats, projects, artifacts, connectors, and skills carry over. Claude asks before taking actions by default, while users can opt for fewer check-ins. For recurring work, a report and its matching slides can be created in one conversation and scheduled to run weekly.

Claude Docs 和 Claude Slides 此次推出，Claude Design 也进入对话；三者均面向付费计划提供 beta 版本，由 Enterprise 管理员控制是否启用。用户可以直接编辑、给幻灯片留评论、在 Claude 中演示、导出 PowerPoint 或 PDF，并通过一个链接分享成果。原有 Cowork 对话、项目、artifacts、connectors 和 skills 会保留。Claude 默认在采取行动前询问，用户也可选择减少确认次数。对于重复性工作，报告及配套幻灯片可以在同一段对话中生成，并安排每周自动执行。

[Claude Blog 原文](https://claude.com/blog/cowork-is-now-claude)