What Claude Cowork Actually Does (And Why It's Different)

TL;DR

Claude Co-work turns LLM interaction into persistent collaboration by working inside shared folders and files rather than ending with a one-off chat response.

Briefing Cornell Notes

Briefing

Claude Co-work’s core shift is turning AI from a chat-only assistant into a persistent, file-based collaborator that can actually execute work inside a shared workspace. Instead of ending with “where did the value go?” Co-work keeps results in documents and folders over time, letting Claude read, edit, and manage tasks alongside the user—an approach Anthropic is positioning as more than a conversation.

The product arrives after Anthropic’s earlier push with Claude Code, which developers began using far beyond software—spreading into research, project management, spreadsheet work, and other computer tasks. That broader pattern helped set up Co-work, launched as a more accessible, user-friendly way to bring Claude Code’s capabilities to non-developers. Co-work runs in a MacOS desktop app with three tabs: Chat (the familiar LLM interface), Code (an advanced mode), and Co-work (the middle tab). The ordering is framed as beginner mode, advanced mode, then “chat plus,” where the AI can operate on real artifacts rather than only responding in text.

What stands out in the interface is the move from “new chat” to “new task,” and from talking about work to “knocking something off your list.” Co-work is also labeled an early research preview/beta, so behavior may be inconsistent. But the workflow is concrete: users can choose “work in a folder,” grant access once (including subfolders), and let Claude treat that folder as its workspace. In the transcript’s example, the user selects a book manuscript folder (“lip”), and Co-work immediately starts analyzing the Word document, showing progress and a private task list as it checks items off.

For a ~45,000-word manuscript, Co-work demonstrates a parallelization strategy: it reads in chunks, uses sub-agents to process different sections simultaneously, then compiles comprehensive feedback. The output isn’t just critique; it includes a prioritized framework (critical issues, high impact, accessibility gaps, structural pacing problems, and specific fixes). The transcript emphasizes that this framework wasn’t provided by the user—it emerged from Claude’s own analysis. It also highlights a practical limitation of standard chat: feeding the entire manuscript at once previously failed, forcing chapter-by-chapter work that reduced editorial quality.

Cost and availability are also part of the pitch. Co-work and Claude Code are described as available on Anthropic’s lowest paid plan (about $20/month at filming time), while earlier access to Claude Code required a higher tier (roughly $150–$200/month). The MacOS app is not yet available for Windows, though Windows support is expected.

A second example shows Co-work’s “plan mode” and web research. With context turned off, Claude researches family-friendly options for a road trip near Vaia de Bravo, Mexico, then asks targeted follow-up questions—like timing (mid-February) and accommodation preferences (a mix). Recommendations evolve as uncertainty is surfaced, such as removing Monarch Butterfly Sanctuaries when the user notes the location is too close to their home to justify the plan. The final deliverable is a day-by-day itinerary exported into a Word document with driving times, morning/afternoon activities, and practical notes like roads, weather, and food—details the user didn’t explicitly request.

Finally, Co-work’s collaboration model has constraints: work is stored locally on the machine, so switching devices won’t preserve conversations. Context is managed through folders rather than uploading files one by one, requiring a mindset shift. The overall message is that Co-work makes LLMs feel less like a “talking buddy” and more like a seated collaborator working from shared documents—more productive, more grounded, and easier to operationalize when results must persist beyond a single chat window.

Cornell Notes

Claude Co-work reframes LLM use from chat into ongoing collaboration by working inside persistent folders and shared files. Instead of producing only a final message, it can read, edit, and manage tasks over time, with progress tracked through a private task list. In a book-editing example, Co-work analyzes a ~45,000-word manuscript by chunking the text and using sub-agents in parallel, then returns prioritized, actionable feedback. In a road-trip planning example, “plan mode” triggers web research and follow-up questions, iterating the itinerary as uncertainty and constraints change. The practical payoff is deliverables that can be exported (e.g., a Word document itinerary) and used offline, though data storage remains local to the device.

How does Claude Co-work differ from standard chat in what “results” look like?

Standard chat typically ends with text output that the user must translate into real work. Co-work instead operates on a shared workspace: users grant access to a folder, and Claude reads and edits files there. The transcript highlights that Co-work persists work over time and can generate artifacts like updated documents and a printable Word itinerary, so value doesn’t disappear when the conversation ends.

Why does “work in a folder” matter for large tasks like editing a book manuscript?

A folder-based workspace lets Claude access a complete set of materials (including subfolders) without manually selecting individual files one by one. In the manuscript example, the user selects a folder containing the latest Word document version. Co-work then treats that folder as its playground, reads the manuscript in chunks, and uses sub-agents to analyze different sections in parallel—something the transcript contrasts with standard chat’s difficulty handling the full context at once.

What evidence suggests Co-work is managing tasks internally rather than just answering prompts?

Co-work shows progress and a private to-do list, checking tasks off as it proceeds. The transcript also notes “thought process” and “context” views that reveal which folder it’s accessing and which files it’s creating or editing. For the book, the interface lets the user click into a work-in-progress area to see a play-by-play of chunking and parallel section analysis.

What is “plan mode,” and how does it change the quality of planning outputs?

Plan mode makes Claude ask critical follow-up questions instead of rushing to a confident answer. In the road-trip example, it asks when the trip is planned (mid-February) and what accommodation style is preferred (a mix). It also surfaces uncertainty and explains what would change recommendations, allowing the user to iterate—such as revising the plan to remove Monarch Butterfly Sanctuaries.

What deliverables does Co-work produce that are useful in real-world constraints?

Co-work can export results into practical formats. The transcript describes Co-work creating a Word document containing a day-by-day itinerary with driving times, morning/afternoon activities, and practical notes (accommodations, roads, weather, food). This is framed as especially useful when cell service is unreliable during travel.

What limitations should users expect regarding storage and device switching?

Work is stored locally on the machine. The transcript warns that switching to a different computer, tablet, or phone won’t carry over prior conversations or files, so users should plan their workflow around where Co-work is installed and where the workspace lives.

Review Questions

In what ways does folder-based context improve Claude Co-work’s ability to handle large, multi-part tasks compared with uploading or pasting content into a chat?
Describe how plan mode changes the interaction flow and why that matters for decision-making under uncertainty.
What practical constraints (storage location, device switching, offline use) influence whether Co-work fits a user’s workflow?

Key Points

1
Claude Co-work turns LLM interaction into persistent collaboration by working inside shared folders and files rather than ending with a one-off chat response.
2
The interface shifts from “new chat” to “new task,” framing the workflow as completing items on a to-do list.
3
Granting folder access once (including subfolders) avoids manual file-by-file selection and enables Claude to explore, research, and edit within a workspace.
4
For large documents, Co-work can chunk content and use sub-agents in parallel, then compile prioritized, actionable feedback.
5
Plan mode triggers follow-up questions and iterates recommendations based on uncertainty, improving planning quality over one-shot answers.
6
Co-work can produce usable deliverables like Word documents for printing and offline reference.
7
Co-work work is stored locally, so switching devices won’t preserve prior conversations or workspace state.

Highlights

Co-work’s “work in a folder” model makes AI output persist as editable artifacts, not just chat text.

A ~45,000-word manuscript was processed by chunking and sub-agents in parallel, producing a prioritized critique with specific fixes.

Plan mode adds follow-up questions and uncertainty-aware iteration, then exports a day-by-day itinerary into a Word document.

Topics

Claude Co-work
Persistent Workspaces
Plan Mode
Parallel Sub-Agents
Local Storage