Meet Cowork

Up to now, everything you have done with Claude has been a conversation. You ask, it answers, you refine. Even when you built it into a workflow in the last module, you were still the one running each step. Cowork changes all of that. Instead of asking for one thing at a time, you hand over a whole task and Claude does the steps itself: reading the files, doing the work in order, and handing back a finished result.

Cowork is accessible on the Home tab of the Claude desktop app, next to the Code tab. The difference from chat is not that it is smarter. It is that it works across steps and directly touches files. You point it at a folder on your computer, describe what you want done, and it reads, edits, and creates files inside that folder until the task is finished. Chat advises you on how to do something. Cowork does it, then shows you what it produced.

Cowork lives in the desktop app, not the browser, so the first step is getting the app onto your machine.

Handing off to Cowork feels a little different the first time. You point Cowork at a folder and describe the outcome you want. Then it works through the steps, and when it is done it hands back the result along with a short account of what it did. You review that result before you use it.

1Point
Aim Cowork at a folder of files.
2Describe
Tell it the outcome you want.
3Review
Check the result before you rely on it.

Here is what a handoff looks like on a basic task that is suited for Cowork. Say you have a folder of PDFs and notes, and you want one clean summary of what is in them.

You

I have a folder called Green Acres month end close with six files in it, a mix of PDFs and notes. Read all of them and write me one short summary of what they cover, saved as a Word doc in the same folder.

Claude

Here is what I did: read all six files, pulled the key points from each, and wrote them into a single one-page summary grouped by topic. I saved it as GreenAcresSummary.docx in that folder and left your originals untouched.

Two things worth a look before you use it: the figures in the second section came from the spreadsheet, so confirm them against the original, and one of the notes was handwritten and partly unreadable, so I flagged that line instead of guessing at it.

Notice what did not appear in that exchange. The actual work, opening each file, reading it, deciding what mattered, writing and formatting the document, all happened on its own between your request and the result. You did not watch it or run any of it. That is the part you handed off.

And notice what did appear. Claude did not just hand back the file and call it finished. It told you what it produced and, more usefully, what to check: a figure that should be confirmed against its source, and a spot where it was unsure and said so instead of inventing an answer. It can be fluent and still wrong, so the summary is a draft until you have checked the two things it flagged and skimmed the rest. The agent did the assembly. You do the checking.

This task needed no skill, which was the point: the act of handing work off is the same whether the task is simple or not. Next you will run a skill the firm built for a real job, splitting a return into one K-1 per recipient, where the stakes are higher and the checking matters more.