Claude Cowork (Fully Tested): I might get ADDICTED to this NEW CLAUDE CODE FEATURE!

Key Concepts

• Anthropic Co-work: A new tool from Anthropic, built on Claude Code, designed for non-developers to interact with Claude for various tasks through a graphical interface.
• Claude Code: Anthropic's agent SDK that allows Claude to perform coding-related tasks and interact with local files.
• Agent SDK: Software Development Kit for building AI agents, providing tools and frameworks for creating autonomous or semi-autonomous AI systems.
• Graphical Interface: A visual way for users to interact with an application, making it more accessible than command-line interfaces.
• File System Access: Co-work's ability to read, edit, and create files within a user-specified folder on a local computer.
• Agency (AI): The capacity of an AI system to act independently and make decisions to complete a given task, often involving planning and execution.
• Connectors: Existing integrations that link Claude to external information sources or services.
• Skills (AI): Pre-defined abilities or functions added to Claude to enhance its performance in specific tasks, such as document or presentation creation.
• Browser Access: The capability for Claude to interact with web content when paired with a browser extension (e.g., Claude in Chrome).
• Parallel Task Execution: The ability to queue and process multiple tasks simultaneously, improving efficiency.
• Access Control: User-defined permissions that limit Claude's access to specific folders and connectors, enhancing security.
• Potentially Destructive Actions: AI actions that could lead to unintended data loss or modification, such as deleting files.
• Prompt Injections: A security vulnerability where malicious input attempts to override or alter an AI agent's intended instructions or behavior.
• Agent Safety: The ongoing field of research and development focused on securing AI agents' real-world actions and preventing unintended or harmful outcomes.
• RPC Error: (Remote Procedure Call error) A technical error indicating a failure in communication between different software components, often related to network or system resource issues.
• Virtual File System: An abstraction layer that allows a program to interact with different file systems in a uniform way, often used for security or isolation.
• LLM (Large Language Model): A type of artificial intelligence model trained on vast amounts of text data, capable of understanding, generating, and processing human language.
• Verdant: An alternative graphical interface tool for interacting with various AI models, highlighted for its fleshed-out UI and flexibility.

Introduction to Anthropic Co-work

Anthropic has launched "Co-work," an extension of its Claude Code capabilities presented within a user-friendly graphical interface. While Claude Code was initially designed for developers and coding tasks, Anthropic observed that users quickly adapted it for a wide range of non-coding activities. This insight led to the development of Co-work, aiming to provide a simpler, more approachable way for anyone, not just developers, to leverage Claude's agency for diverse work.

Differentiating Co-work from Standard AI Conversations

Co-work fundamentally differs from a regular conversational AI experience by granting Claude direct access to a user-selected folder on their computer. This access allows Claude to read, edit, or create files within that specific directory. Unlike a typical chat where users manually provide context and convert outputs, Co-work enables Claude to operate with significantly more agency. Once a task is assigned, Claude formulates a plan and systematically executes it, providing updates on its progress. This operational model is built upon the same foundations as Claude Code, utilizing the Claude Code agent SDK, making it capable of handling many of the same tasks as Claude Code but in a more accessible format for non-coding workflows.

Specific examples of Co-work's capabilities include:

• Reorganizing downloads by sorting and renaming files.
• Creating new spreadsheets from a collection of expense screenshots.
• Producing a first draft of a report from scattered notes.

Advanced Capabilities and Customization

Co-work offers several advanced features to enhance its utility:

• Existing Connectors: Claude can integrate with pre-configured connectors that link it to external information sources, expanding its knowledge base.
• Initial Set of Skills: Anthropic has incorporated specific skills to improve Claude's proficiency in generating various file types, such as documents and presentations.
• Browser Access: When paired with "Claude in Chrome," Co-work can perform tasks that necessitate web browsing, further extending its operational scope.

User Experience and Workflow Design

Co-work is engineered for simplicity and efficiency in managing new work. It eliminates the need for users to constantly provide manual context or convert Claude's outputs into desired formats. Users can queue up multiple tasks, allowing Claude to process them in parallel. This design shifts the interaction model from a back-and-forth conversation to a more asynchronous, collaborative dynamic, akin to "leaving messages for a coworker."

Safety Considerations and Risk Mitigation

While Co-work offers powerful capabilities, Anthropic emphasizes the importance of user awareness regarding potential risks:

• Access Control: Users maintain explicit control over which folders and connectors Claude can access, ensuring it cannot read or edit unauthorized content.
• Action Confirmation: Claude is designed to request user confirmation before executing any significant actions, providing opportunities for steering or course correction.
• Potentially Destructive Actions: Users must be aware that Claude, if instructed, can perform actions like deleting local files. Clear and precise instructions are crucial to prevent misinterpretation and unintended data loss.
• Prompt Injections: There is a risk of attackers attempting to manipulate Claude's plans through malicious content encountered online. Anthropic has implemented "sophisticated defenses" against prompt injections, but acknowledges that "agent safety" – securing Claude's real-world actions – remains an active area of industry development. Users are advised to take precautions, especially while familiarizing themselves with the tool.

Practical Demonstration and Use Cases

The presenter demonstrated Co-work's functionality through several real-world tests:

Setting Up and Initial Challenges

Accessing Co-work involves opening the Claude desktop app and selecting the Co-work option. Users then input their prompt and select the target folder. During an initial attempt to organize files, the presenter encountered issues:

• A loading spinner persisted for 30 minutes without a response.
• A buggy thread displayed two messages sent despite only one prompt being entered.
• An "RPC error failed to mount vault" occurred, suggesting Co-work might use a virtual file system that copies data for security, which can be "finicky." After restarting and a brief delay, the tool began working.

Use Case 1: File Organization and Content Analysis

• Task: Organize approximately 1,200 markdown script files (each ~10,000 characters) from an Obsidian vault, specifically moving all scripts about "GLM" into a new folder named "GLM," making copies rather than moving originals. The challenge was that simple title matching wouldn't work due to "weird naming schemes," requiring content analysis.
• Process: Claude initially asked for clarification on "what GLM is" (which the presenter found a "dumb question" given it's an LLM). After clarification, it proceeded efficiently.
• Result: Claude successfully organized the scripts into the designated folder, demonstrating its ability to check file contents for accurate categorization. The UI also allowed viewing "added context" and previews within the session.

Use Case 2: Report Generation and Data Analysis

• Task: Generate a "rewind report of 2025" based on the script files.
• Process: Claude quickly created tasks, analyzed the scripts, and produced a document file. The process was described as "very snappy."
• Report Details: The generated report was comprehensive, including:
  • Total number of scripts in 2025.
  • Total words and average words per script.
  • A monthly breakdown, identifying March as the busiest month (93 scripts) and December as the lightest (43 scripts), which aligned with model release cycles.
  • An analysis of covered topics, highlighting "AI tools," "Gemini," "Chinese models," "CLI tools," and "Claude."
  • Identification of the longest scripts and key insights.
• Evaluation: The presenter found the report "kind of awesome" and "a really good one for sure."

Use Case 3: Prototype Creation (Coding)

• Capability: Co-work can "make prototypes" by coding and generating artifacts.
• Connection: This feature leans into the "Claude Code section" but makes it more accessible, addressing the "barrier for entry for newer people" who might find Claude Code itself daunting.

Critique, Areas for Improvement, and Future Outlook

The presenter offered several points of critique and suggestions for future development:

• Error Handling: Technical errors like the "RPC error" are not user-friendly for non-developers, suggesting a need for improved and more accessible error messages.
• Integration Potential: The presenter speculated on future integrations, such as "Claude Co-work on the web" or integration with messaging platforms like WhatsApp or Discord, similar to "Claudebot."
• Comparison to Alternatives: The presenter mentioned "Verdant" as a strong alternative, praising its "very graphical interface," "super fleshed out UI," and flexibility in changing system prompts to act as a simple AI agent. While Co-work is "good to have," Verdant is still recommended for its robust features. Overall, Co-work is considered "pretty cool" and a valuable addition to Anthropic's offerings.

Conclusion and Main Takeaways

Anthropic's Co-work is a significant step towards making advanced AI agent capabilities accessible to a broader audience beyond developers. By providing a graphical interface and direct file system access, it empowers users to automate and streamline various non-coding tasks, from file organization and report generation to prototype creation. While it offers impressive agency and efficiency, users must exercise caution regarding security risks like prompt injections and potential destructive actions, providing clear instructions. Improvements in error handling and potential integrations with web and messaging platforms could further enhance its utility and user experience. Despite some initial bugs and the existence of strong alternatives like Verdant, Co-work represents a powerful and promising tool for collaborative work with AI.