what claude cowork actually is.

elisabeth hitz · june 18, 2026 · 6 min read

there are now three ways to work with claude on the desktop, and most people only know one of them. chat, cowork, and code. they look similar and they are built for completely different shapes of work. getting this distinction right is the difference between asking AI about your work and handing it the work.

this is about cowork, the middle one, and the one most owners will get the most out of.

chat asks, cowork does

chat is a conversation. you ask, draft, brainstorm, think out loud, and claude answers in the window. it is turn by turn. it can pull from a tool, but you stitch the steps together yourself. it is built for bounded exchanges and quick thinking, and it is still where you go to work a problem out.

cowork is a workspace. you point claude at a folder on your computer, connect the apps your work lives in, and describe an outcome. claude makes a plan, works through the steps, uses the tools and files it needs, and saves a real deliverable back to your folder.

the shift is from chatting to delegating. most people come from chat and keep typing a question, reading the answer, typing a follow-up. that works in cowork too. but you get the most out of it when you reach for the work you would normally do yourself, not the work you would normally ask about.

what makes a good cowork task

the shape is specific. a cowork task usually has all four of these:

• it touches local files. the inputs or the output live in a folder on your machine.
• it spans more than one tool. email, calendar, drive, a spreadsheet, a slack thread.
• it takes several steps. gather, read, analyze, assemble.
• it ends in a real deliverable. a doc, a deck, a sheet you can send.

a clean example: pull together a one-page brief on last quarter's results, where the numbers are in a spreadsheet on your desktop and you want the brief saved to your reports folder. claude reads the sheet directly, so the numbers are accurate with no copy-paste, and lands the file where you asked. the presence of a real file and a real save location is the signal that it is cowork's kind of work, not chat's.

it works in your environment, not next to it

the reason this matters is that cowork operates where the work already is. on your files, it reads the folder you point it at and writes finished output back. in your apps, it pulls context from the email, calendar, messaging, and drive you have connected. in your browser, for web tools without a connector, it can read and act on the page itself. and it takes action, it does not just describe what to do.

that is the same connected-tools idea i wrote about in connect claude to your tools, applied to a whole piece of work instead of a single question.

you stay in control

the obvious worry with a thing that takes actions is exactly the right worry, and cowork is built around it. claude shows you its plan before it starts. by default it asks before anything that matters, sending, deleting, sharing. and you can steer at any point.

it is the same draft-and-approve model that makes a claude setup safe in the first place: claude does the reading, sorting, and drafting, and a human approves the things with consequences. i wrote about why that model is the whole game in how to set up claude for a small business.

chat, cowork, or code

quick rule. chat is for thinking with claude: questions, brainstorming, a draft you will polish yourself. cowork is for delegating to claude: multi-step work across files and tools that ends in a deliverable you set in motion and come back to. code is for building software with claude: working inside a codebase, editing files, running tests, making commits, developer work, not document work.

most owners will live in chat and cowork and use them for very different things. the move is to stop reaching for chat out of habit when the task is really a piece of work you would have done by hand.