Claude Cowork: What It Is and How to Use It Well

Claude Cowork is where you hand Claude real work instead of chatting about it. You describe an outcome, choose the folders and tools it may use, and Claude plans the steps, does them and delivers finished files for your review. It runs on the same agentic architecture as Claude Code, with no terminal required.

I teach teams to use Claude every week, and Cowork is the feature that makes non-technical people say "oh, now I get it." It's also the one that disappoints fastest when it's set up badly. Here's what it does, what it costs and the setup I recommend.

What is Claude Cowork?#

According to the Claude Help Center, Cowork brings Claude Code's agentic capabilities to knowledge work beyond coding. Instead of answering prompts one at a time, Claude takes on complex, multi-step tasks and executes them on your behalf.

When you start a task, Claude:

1. Analyzes the request and creates a plan.
2. Breaks complex work into subtasks.
3. Runs code and commands in an isolated environment on Anthropic's servers.
4. Coordinates parallel workstreams when that helps.
5. Delivers finished outputs you can preview and download.

It works with the files people actually use: Word, PDF, Excel with working formulas, PowerPoint, CSV, Markdown and images. Sessions run in the cloud, so work continues when you close your laptop, and you can check in from desktop, web or mobile.

Claude Cowork vs Chat vs Claude Code#

Anthropic's own product FAQ draws the line clearly: Claude Code is for writing, debugging and shipping code; Cowork is for research, analysis, document creation and similar tasks. If you're choosing a tool for developers, start with how to use Claude Code instead.

How to use Claude Cowork, step by step#

1. Start a task#

Open Claude on the web, desktop or mobile, select "Cowork" in the message box and describe the task. If the task needs local files, your browser or other apps on your computer, keep the Claude Desktop app open.

2. Write the outcome, not the steps#

The best Cowork prompts read like a brief to a capable colleague. Anthropic's example for a monthly close is a good model: where the files are, what to compare, the threshold that matters, the tabs you want and what to flag. Say what "done" looks like and what to do when something doesn't reconcile.

3. Choose an approval mode#

Cowork has three modes, selectable from the chat box:

• Manual: Claude asks before each action. Use it for new tasks and anything that sends messages.
• Auto: Claude reviews each action for risks such as prompt injection or data exfiltration and blocks unsafe ones. It uses more of your limits.
• Skip: no pauses and no automatic checks. Only for tasks where you trust every file, connector and site involved.

4. Review and steer#

Claude shows the files it opens, the tools it uses and the choices it makes. Jump in when it heads the wrong way, and edit Markdown drafts in place by highlighting text and asking for a change.

The setup that makes Cowork dependable#

Most bad Cowork results come from missing context, not a weak model. Three settings fix most of it:

• Global instructions in Settings, then Cowork: your role, preferred formats, tone and the things Claude should never do.
• Folder instructions for each project folder: what the files are, naming conventions and where outputs go. Claude can update these as it learns.
• Plugins and skills: plugins bundle skills, connectors and sub-agents so Claude shows up as a specialist. Your own procedures belong in skills, which I explain in Claude skills.

Then use /schedule for anything recurring. Scheduled tasks run in the cloud on a cadence you choose, so the Monday report is waiting instead of being assembled.

What does Claude Cowork cost?#

Cowork is included in paid plans, per Anthropic's pricing on the product page:

Agentic work consumes limits faster than chat. If you hit them, batch related work into one session and keep simple questions in regular chat. For the wider picture of Claude plans, see Claude Code pricing.

Current limitations#

• No session sharing. You can share artifacts from a session, but not the session itself.
• Some features need the desktop app, including plugins with local MCP servers and access to local files.
• More usage per task than chat, and more again in Auto mode.

Use cases that work on day one#

Good first tasks share three traits: the inputs already exist, the output has a clear shape, and a mistake is cheap to catch.

• Turn a folder of receipts into a formatted expense report.
• Summarize a quarter of call notes into objections, quoted from each note.
• Build a weekly metrics deck from exports in a shared drive.
• Rename and sort a messy downloads folder.
• Draft a first-pass review of contracts against your playbook, with clause citations.

Save the risky ones, like sending emails or changing records, until your instructions have held up for a few weeks. If you want your whole team using Cowork and Claude on their real work, that's what my team training is built for, and how to build an AI agent covers the same handoff principles for more technical setups.