Claude Cowork — unicorn

In short Cowork is a mode where Claude works directly with your files, folders and apps: reading, editing and producing real results on your machine. If Chat is a conversation, Cowork is a working session: you describe a multi-step task, Claude plans and executes, you steer and accept.

A working session: you set and accept, the agent plans and does.

What it is and how it differs

Its place among the three ways to work is between "conversation" (Chat) and "code" (Code). Cowork is for knowledge workers: assemble a report from a dozen files, triage the mail and draft replies, make a deck from data. Not a turn-by-turn dialogue but a handed-over task with a finished result.

Analogy: a capable assistant. You set the goal and check the outcome rather than dictating every keystroke. 🔴 Needs a paid plan (Pro/Max/Team/Enterprise) and the desktop app. No code required.

Setting it up

A working folder — where Claude reads and creates files; it acts within its bounds.
Connectors — access to the apps you need (connectors, Enterprise Search and Research).
Permission mode — how much freedom: ask before actions, or act on its own. The same supervision logic as in Claude Code.

How a task goes

Describe the task (the more precise, the better, prompting and AI Fluency).
Claude asks clarifying questions — answer them.
Claude plans and executes multi-step (the agent loop under the hood).
You steer along the way — step in and correct.
Accept the result — check the finished file.

Making Cowork yours

Standing context — global instructions and projects (Claude projects and artifacts).
Skills — your recurring processes (Claude skills).
Plugins — bundle skills and share with a team through the org marketplace.
Where else it works: Claude in Chrome (web apps) and Claude for Microsoft 365 (Word, Excel, PowerPoint, Outlook).

Safety is mandatory

Since Claude acts on your behalf with real files, the responsibility is yours (the Diligence habit): check the plan before it runs and the result after; start in approval mode; validate skills (evals) before relying on or sharing them; understand the bounds of the working folder and connector access.

The source is the Introduction to Claude Cowork course. Related: three ways to work with Claude, Claude projects and artifacts, Claude skills, the agent loop and tools, prompting and AI Fluency.