X / Twitter
Boris Cherny, Anthropic Claude Code Team
Boris Cherny announced that Claude's chat and Cowork are beginning to merge, with the aim of letting the same Claude carry context across different tasks. He used Claude Code and Cowork to illustrate the foundation of this change: developers hand over feature requirements and receive code, while knowledge workers hand over task descriptions and receive finished files. AI's value is shifting from answering questions to completing work. Claude Docs, Claude Slides, and Claude Design are also coming into the conversation, allowing users to ask directly for documents, presentations, or designs without finding a separate tool entry point. Generated presentations can be opened, edited, and exported as PowerPoint or PDF. He said he had been using the experience for several weeks and found it simpler, faster, and more capable. The new experience will roll out gradually, with the team continuing to improve speed and reliability along the way.
https://x.com/bcherny/status/2100260544087535639
https://x.com/bcherny/status/2100259951398789487
Peter Yang, AI Tutorial and Interview Creator
Peter Yang shared a workflow for producing an entire podcast with 8 AI skills, while emphasizing that he remains involved in aesthetic judgments and editorial decisions. During preparation, /podcast-prep researches the guest and generates an interview outline. During editing, /podcast-edit reviews the raw transcript, then works with him to select standout quotes for the opening clip and passages to remove. During production, /podcast-production orchestrates another 5 skills to turn the complete content into 6 different assets. He rejected the idea that "the latest models have made skills redundant," because these skills make AI follow his specific editing requirements and browser-operation instructions. He also released a video demonstrating all 8 skills in use, as a reference for people looking to automate manual workflows.
https://x.com/petergyang/status/2100328939034128856
Madhu Guru, Senior Director at Meta AI
Madhu Guru argued that safety and security are themselves features of AI products and models. He discussed both capabilities at the product and model levels. His central point was that safety should be a capability the product itself provides, as opposed to being understood as an externally imposed constraint. With this framing, he reminded AI builders to include safety capabilities when discussing product features.
https://x.com/realmadhuguru/status/2100312717739667963
Cat Wu, Anthropic Claude Code and Cowork Teams
Cat Wu explained the rationale for merging Cowork and chat through user feedback: many people do not want to decide which Claude product to use every time they begin a task. As model capabilities improve, the team believes Claude can now route tasks itself. Based on the prompt, it will decide whether to give a quick answer or perform deeper agent work and choose the right output format for the task. Claude Design is also integrated, so creating slides, designs, or documents no longer requires switching tabs or applications. Automatic routing still preserves user control: users can stop a task, change direction, or exercise finer control over how much effort Claude puts in and how it executes at any time. The team plans to roll out the experience gradually over the next few weeks and refine it based on feedback.
https://x.com/_catwu/status/2100260655312089562
Thariq, Anthropic Claude Code Team
Thariq updated his view on agent tool design: it is increasingly possible to give Claude tools that directly match real needs, without an intermediate wrapper to guide it. Using data storage as an example, he suggested first deciding whether what one actually needs is a file system or an API that directly accesses a database. He explicitly said this differs from his understanding last year. For runtime architecture, he thinks Claude Managed Agents strikes a good balance because the sandbox is optional and can be started independently of the rest of the agent's execution loop. He recently migrated an older project centered on bash tool calls to Claude Managed Agents and reported that it works well. Together, the two posts emphasize choosing tool interfaces and execution environments according to the task's needs.
https://x.com/trq212/status/2100315538472009897
https://x.com/trq212/status/2100315537251463523
Guillermo Rauch, Vercel CEO
Guillermo Rauch shared results from his team's use of Typesafe AI, focusing on security reviews before commands are executed. In 𝚏𝚡, the default auto mode has a security reviewer analyze every command. That reviewer currently runs on GPT Luna. He said Jev delivered up to an 18-fold speedup at p95, along with greater accuracy. Jev will be added to Vercel AI Gateway and may become the new default, though the post did not present the default switch as a settled decision.
https://x.com/rauchg/status/2100307962262872105
Aaron Levie, Box CEO
Aaron Levie focused on a previously underappreciated category of AI innovation: highly capable information processing at extremely low cost and very high speed. He believes this combination has broad value for enterprise tasks, particularly the recurring decision points found across many processes. Examples he listed included data classification, routing decisions within workflows, and judgments about domain-specific questions. Quickly assessing safety or security issues was another application he named. These judgments often determine whether a process can advance to its next step, so the cost and speed of handling them affect the entire workflow. He believes these models and approaches deserve attention in enterprise agent workflows.
https://x.com/levie/status/2100448648672993540
Zara Zhang, Product Builder
Zara Zhang said she had recently been using Claude less because she found its way of expressing itself increasingly hard to tolerate. Her criticism focused on the attitude conveyed by Claude's speech. She felt it was constantly showing off how clever and profound it was, while clearly communicating a point was becoming less of a priority. This feedback directly connected the model's expressive style with her willingness to use it and made clear that her dissatisfaction with Claude centers on the communication experience.
https://x.com/zarazhangrui/status/2100278750776824115
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari has spent months refining a household assistant called "Home" using Nous Research's agent and said it had recently reached a very good level of usability. It is also the first group bot that he and his wife both use every day, with their shared interactions turning failures into learning opportunities. He tried to reproduce it with tools such as Grok Bot but felt they could not provide the same degree of granular control. Specific requirements he listed included reading only designated emails, converting attachments embedded in emails into structured data, and using an already authenticated browser session. These examples show that his concerns go beyond conversational ability to include control over what is read, how data is processed, and browser state. He also asked the agent to outline its own responsibilities based on its existing capabilities and conversation history, but the supplied material did not include the details of that response.
https://x.com/nikunj/status/2100212813625196917
Claude, Anthropic's AI Assistant
Claude's official account added details on how to create documents, presentations, and designs within a conversation. Users can first draft a one-page brief with Claude Docs, turn it into a presentation with Claude Slides, and create accompanying visuals through Claude Design, all in the same place. When revising, they can directly edit a sentence in a document, leave a comment on a slide for Claude to address, or select and move an element. Once finished, users can present their slides directly in Claude, export documents as needed, or share everything through a single link. Existing Cowork users will retain their chats, projects, artifacts, connectors, and skills and can pick up where they left off. Claude Design, Claude Slides, and Claude Docs are currently available in beta on all paid plans.
https://x.com/claudeai/status/2100258495543071016
https://x.com/claudeai/status/2100258494221812123
https://x.com/claudeai/status/2100258492590207079
Official Blog
Claude Begins Merging Cowork and Chat
Claude announced the merger of Cowork and chat, bringing quick questions and more complex work deliverables into the same conversational starting point. The official explanation was that Cowork had previously handled larger tasks and Design had handled visual work, but users had to decide where a task belonged and also faced a lack of context continuity between the different entry points. After the merger, Claude will determine what a task needs and execute it using existing context, skills, and connectors; the company also said tasks can keep running after users close their laptops. The integration will first roll out gradually to Pro and Max users over the coming weeks across web, desktop, and mobile, with no manual activation required. Team and Free plans will follow, while Enterprise administrators will receive at least 30 days' notice before their organization's experience changes.
The simultaneously released Claude Docs and Claude Slides, alongside Claude Design's integration into conversations, bring documents, presentations, and visual design into a single workflow. The official example was a weekly report: a user can ask Claude to check changes in the previous week's sales pipeline, flag delayed items, and organize the highlights into 5 slides for a leadership meeting. Because the report and slides come from the same conversation, their content can stay consistent. Users can check progress from their phones, edit content directly, leave comments, and download PowerPoint or PDF files; they can also schedule the task to start automatically every Monday. By default, Claude asks before taking action, but users can choose to let it keep working and request confirmation only when something needs closer review. All three content creation features are in beta on paid plans, with Enterprise administrators deciding when to enable them. The standalone Claude Design continues working as before.