Claude Cowork: Stop chatting, hand off a task

Most people who try Cowork for the first time make the same mistake: they ask it a question.

The question gets an answer, the answer looks fine, and they go back to chat and say “it’s the same thing.” It’s not the same. They just didn’t give it a task.

Cowork is the same agent architecture as Claude Code - just without a terminal. Instead of one answer per prompt, you describe an output and it runs until there’s a file in hand. This guide shows when it beats chat, how it works under the hood, and how to start without burning your usage quota in three days.

The difference between chat and Cowork is who’s at the wheel

In chat, you run the conversation. Question, answer, question, answer. You keep the context in your head, you decide the next step, and you stitch the result together at the end from ten messages.

In Cowork, you describe an output. Claude builds a plan, runs code, works on files, and returns a finished file. You approve at checkpoints instead of sitting at the wheel.

That’s also the cost difference: a Cowork task consumes much more quota than a chat. So “chat or Cowork” isn’t a matter of taste.

Four questions that tell you if this is even a Cowork task

Before you hand something off, check there’s a real task there. These four questions cover most cases:

Does the task touch files - local or in Drive?

Is there more than one step? (read, process, build)

Does the result need to be a file - Excel, deck, document?

Will the work take more than a minute of continuous running?

Simple rule: two “yes” or more - hand it off. Less than that, a chat will get you an answer faster and cheaper.

Three minutes to your first task

Getting started is shorter than it looks:

Paid plan - Pro, Max, Team, or Enterprise.
Desktop app - macOS or Windows. Web and mobile are still in beta.
In the message box, pick Cowork instead of Chat - and describe a task. Not a question. A task.

Step three is the whole difference. “What should I do with these receipts?” is a question. “Take the 40 receipts in the folder and build an expense report” is a task.

How a task actually runs under the hood

Cowork doesn’t send your prompt to a model and return text. It decomposes, plans, runs code in an isolated environment, and spins up sub-agents in parallel. Worth seeing it step by step:

"Take the 40 receipts in the folder and build an expense report"

Analyzis and plan
Breakdown into sub-tasks
Code in an isolated environment
Sub-agents in parallel
The output waits in the session

Two points to notice: the code runs in an isolated environment on Anthropic’s servers, not on your machine. And the session keeps running even if you close your laptop mid-task.

The three approval modes: how much control you let go

Every task has an approval mode, which sets how often Claude stops and waits for you. It’s the most important setting in Cowork, and the one most people skip:

// what happens

Claude reviews each action for safety and decides on its own whether to run it. Risky actions still stop with you.

// when

Everyday work. The sensible default for most tasks.

// the cost

Uses more quota - the checking itself costs.

One permanent exception: deleting files always requires explicit approval. In every mode.

My recommendation for your first task: Manual, on a non-critical folder. It’s slow and annoying, and that’s the point - you see how it thinks before you let it run on its own.

What turns this from a demo into a work tool

Cowork out of the box is a demo. What makes it a work tool is the context you inject:

Global instructions - in Settings. Your tone, your tools, what’s allowed without asking. Loaded into every task.
Folder instructions - instructions that live in a local folder and apply to all work there.
Connectors - Slack, Drive, Jira. Claude reads and acts in the real sources so you don’t copy-paste.
Skills - a packaged procedure you trigger with /skill-name. Or simply: “Package what we just did into a skill.” I wrote more about it in the post on website optimization with Skills.
Plugins - a bundle of skills, connectors, and sub-agents. Install once.
Projects - a workspace with files, instructions, and memory that persists across sessions.
/schedule - scheduled tasks that run in the cloud even when your computer is closed. A task that needs local files will run only when the computer is open. The Sunday report writes itself.

The order here isn’t accidental. Global instructions before connectors, connectors before skills. No point packaging a procedure before it has context to work against.

The prompt builder: goal, source, output, cadence

A good Cowork prompt has four parts: the goal, the source, the output, and the cadence. Build one and copy it:

// goal

// source

// output

// frequency

Build an expense report from the local folder. Go over all the material, build a plan before you start, and the final output: an Excel file with formulas.

Notice what’s missing: an explanation of how to do it. You define the output and the boundaries, not the method.

What people actually hand off to Cowork

Theory is nice, but the convincing part is real tasks. Every card here is a copy-paste-ready prompt:

// Files

Order in the downloads folder

Go over my downloads folder, group files by type and project, and propose a new folder structure before you move anything.

// Files

Expense report from receipts

Take all the receipts in this folder and build an expense report in Excel: a row per receipt, a category, and a monthly summary with formulas.

// Research

Vendor comparison

Research 5 managed CI/CD vendors, compare pricing, limits and SLA, and summarize in a table with a reasoned recommendation.

// Research

Competitor scan

Go over the pricing pages of the competitors on the list, document what changed from what we knew, and prepare a summary of the gap against our own price list.

// Documents

A deck from a written summary

Take the quarterly summary document in the folder and turn it into an 8-slide deck: a slide per topic, numbers big, no walls of text.

// Documents

A README from the code

Go over the repo in the folder and write a README: what the project does, how to run it, the folder structure, and one working usage example.

// Data

CSV cleanup

There is a CSV with customer data in the folder. Remove duplicates, unify phone and date formats, and output a clean file plus a summary of what was fixed.

// Data

A dashboard from logs

Analyze the log files in the folder, find the 10 most common errors and trends over time, and build an Excel report with charts.

The common thread: defined output, defined source, and an explicit request to plan before starting.

Limits worth knowing before you get excited

There are a few things Cowork doesn’t do - good to know upfront, not mid-task:

Quota. Cowork consumes much more usage quota than chat. Long tasks, accordingly.
Memory. Chat memory doesn’t carry over to Cowork. Shared context exists only inside Projects.
Sharing. You can’t share a session with others.
Platform. Some capabilities are desktop-only.
Risk. An agent with access to files and the internet isn’t a chat. Start in Manual on a non-critical folder, and raise permissions only after you understand how it works.

The last item is the only one that can actually cost you. The rest is inconvenience.

First-week checklist

Seven steps, about a week, and after that it’s routine, not an experiment:

0/7

FAQ

What’s the difference between Cowork and Claude chat?

In chat, you run the conversation: question, answer, and you assemble the result at the end from ten messages. In Cowork, you describe an output and Claude builds a plan, runs code, works on files, and returns a ready file. It’s the same agent architecture as Claude Code, without the terminal.

What do I need to start using Cowork?

A paid plan (Pro, Max, Team, or Enterprise) and the desktop app for macOS or Windows. Web and mobile are still in beta. In the message box, choose Cowork instead of Chat and describe a task, not a question.

Does Cowork’s code run on my computer?

No. The code runs in an isolated environment on Anthropic’s servers, not on your machine. The session keeps running even if you close your laptop mid-task.

How much quota does Cowork consume?

Much more than chat, depending on task length. That’s why choosing between chat and Cowork isn’t a matter of taste: if there isn’t a multi-step task with a defined output, a chat will get you an answer faster and cheaper.

Does chat memory carry over to Cowork?

No. Chat memory doesn’t carry over, and shared context across sessions exists only within Projects.

Which approval mode should I choose for my first task?

Manual, on a non-critical folder. It’s slow and annoying, and that’s the point: you see how it thinks before you let it run on its own.

Bottom line

Cowork doesn’t replace chat. It replaces the part where you paste files into a conversation, ask ten questions in a row, and assemble the result yourself at the end.

The moment you spot that happening - stop chatting and hand off a task.

If you want your team to get it without climbing the whole learning curve alone, I give a talk on this in organizations. Talk to me and we’ll see what fits.