I got a private lesson on Claude Cowork & Claude Code

Claude Co-work & Code: A Deep Dive with Boris – Summary

Key Concepts:

• Claude Code: Anthropic’s powerful coding assistant, initially terminal-based, now accessible through various interfaces.
• Claude Co-work: A new UI designed to make Claude Code’s capabilities accessible to a wider audience, focusing on file interaction and tool use.
• Quad Agent: The underlying agentic model powering both Claude Code and Co-work, enabling tool use and computer interaction.
• Agentic AI: AI capable of taking action and interacting with the world beyond text-based responses, utilizing tools and accessing computer resources.
• CloudMD: A text file used to provide Claude with specific instructions and corrections, improving its performance over time.
• Plan Mode vs. Auto-Accept Edits: Two modes within Claude Code/Co-work; Plan Mode for outlining and refining tasks, Auto-Accept Edits for automated execution.
• Skills: Reusable, pre-packaged functionalities within Co-work that extend its capabilities to specific tasks and software.

1. Introduction & The Rise of Agentic AI

The discussion centers around Claude Co-work, a new product from Anthropic designed to simplify access to the power of Claude Code. Boris, the creator of Claude Code and co-creator of Co-work, emphasizes that the initial intention for Claude Code wasn’t a product, but its unexpected versatility led to its development. He highlights a key observation: users quickly moved beyond coding, applying the tool to a wide range of tasks. This mirrors the unpredictable evolution of app stores, where initial simple apps paved the way for groundbreaking applications like Uber and TikTok. The current phase with Claude Code and Co-work is characterized by collective learning and experimentation, with even the creators acknowledging they don’t have all the answers.

2. Understanding Claude Co-work: Functionality & Access

Claude Co-work is currently available for macOS (Windows coming soon) and operates as a desktop application with three main tabs: Chat, Co-work, and Code (which is essentially Claude Code). Crucially, Co-work is built on top of Claude Code, utilizing the same “Quad Agent” SDK. This agent is what enables the tool to take action, interact with files, and use tools on the user’s computer – a defining characteristic of “agentic AI.” Boris clarifies that true agentic AI isn’t just about chatbots or web searching; it’s about the ability to do things. Anthropic prioritized coding and tool use from the beginning, leading to the current capabilities.

3. Core Workflow: File Interaction & Automation

The demonstration focuses on a practical example: renaming receipts in a folder. Co-work requires explicit permission to access specific folders, prioritizing user safety. A key concept is that Co-work operates with the user’s files, not just on copies. The process showcases Co-work’s ability to rename files based on their content. This leads to a discussion of Co-work’s broader capabilities: generating files (presentations, spreadsheets), interacting with tools via MCP (Message Communication Protocol), and controlling web browsers. The initial recommendation is to start by mounting a folder and experimenting with basic tasks like file organization.

4. Safety & Technical Foundations

A significant portion of the discussion addresses the safety measures built into Co-work. These include:

• Model Alignment: Extensive research into the model’s internal workings (akin to studying neurons) to ensure it behaves as intended.
• Virtual Machine: A sandboxed environment to prevent unintended consequences like file deletion.
• Deletion Protection: Prompts to confirm potentially destructive actions.
• Prompt Injection Protections: Measures to prevent malicious inputs from hijacking the system.

These safeguards are not perfect, but Anthropic emphasizes the importance of real-world testing to identify and address vulnerabilities.

5. From Receipts to Spreadsheets & Beyond

The demo progresses to converting the receipt data into a spreadsheet, then to a Google Sheet, demonstrating Co-work’s ability to adapt to different formats and applications. This highlights the tool’s potential as a “teammate” capable of handling complex tasks with minimal instruction. The ability to automate tasks like sending the spreadsheet to a finance team via email is also showcased, demonstrating integration with email and Slack.

6. The Power of Parallelism & The "Tend to Your Clouds" Workflow

Boris advocates for a workflow of running multiple “clouds” (instances of the agent) in parallel. He describes a process of initiating tasks in separate tabs, allowing the agent to plan and execute them concurrently. This approach, he argues, is more efficient than focusing on one task at a time, especially with the improved capabilities of Opus 4.5. He describes his workflow as “tending to his clouds,” checking on their progress and providing guidance as needed.

7. ClaudeMD: The Knowledge Base for Improvement

The concept of the “CloudMD” file is introduced as a crucial element for improving Co-work’s performance. This is a simple text file that users can edit to provide feedback and correct errors. By adding specific instructions and examples to the CloudMD, users can teach Co-work to avoid repeating mistakes. The team at Anthropic uses a shared CloudMD for the Claude Code repository, contributing to a collective knowledge base. This is likened to a lint rule in traditional software development, automating the correction of common errors.

8. Opus 4.5 & The Importance of Verification

Boris strongly recommends using Opus 4.5 with “thinking” enabled, as it delivers superior results and efficiency compared to smaller models. He emphasizes the importance of giving Co-work a way to verify its output, such as using the Chrome extension to test code or preview results in a browser. He draws an analogy to a painter working blindfolded, highlighting the need for feedback to achieve high-quality results.

9. Future Outlook & The Exponential Growth of AI

Looking ahead to January 2027, Boris acknowledges the difficulty of predicting the future of AI, given its exponential growth. He predicts that Co-work will become increasingly capable of automating tedious tasks, freeing up users to focus on more creative and strategic work. He envisions a future where users can hire “digital teammates” with specialized skills, similar to vertical job boards. He also notes the potential for Co-work to become a gateway to more advanced Claude Code features.

10. Final Thoughts & The "Clo" Debate

The conversation concludes with a discussion of the viral tweet showcasing Boris’s Co-work setup and a lighthearted debate about whether to call Claude “Clo” in French Canada. Boris agrees that “Clo” sounds better and promises to test it out with his team.

Notable Quotes:

• “The biggest difference with agents is it can take action and you know it's not just text and it's not just like web searching but it can actually use tools on your computer.” – Boris
• “Once the plan is good, the code is good.” – Boris
• “It’s sort of like a find your own path book… there’s no one right way to do it.” – Boris
• “You don’t have to point anything out twice.” – Boris (referring to the CloudMD)

Technical Terms:

• MCP (Message Communication Protocol): A protocol for tools to communicate with AI agents.
• Quad Agent: The underlying agentic model powering Claude Code and Co-work.
• Agentic AI: AI capable of taking action and interacting with the world.
• CloudMD: A text file used to provide feedback and improve Claude’s performance.
• Opus 4.5: Anthropic’s current large language model, known for its coding and reasoning abilities.
• Skills: Reusable functionalities within Co-work.
• Plan Mode: A mode for outlining and refining tasks before execution.
• Auto-Accept Edits: A mode for automated execution of tasks.
• Prompt Injection: A security vulnerability where malicious inputs can manipulate the AI’s behavior.

This summary aims to provide a detailed and specific overview of the conversation, preserving the technical precision and nuances of the original transcript.