Cowork is Anthropic's agentic mode for knowledge work. You describe an outcome, Claude works through the steps, and you come back to finished output rather than a reply. Chat is a conversation where you are in every turn. Both run on the same models, and the difference is how much of the work happens without you. Cowork is available on paid Claude plans, in the same place you already chat, by selecting Cowork in the message box.
The mental shift is smaller than the feature list makes it look, and it is the whole thing.
Chat is asking. Cowork is delegating.
In chat you are present for every step, which is exactly what you want when the shape of the work is still unclear. You are thinking out loud and correcting as you go, and the value is in the back and forth.
In Cowork you specify up front and step away. Claude plans, breaks the work into parts, runs them, and returns something finished. You review at the end rather than steering at every turn.
Which means the skill that separates a good Cowork result from a poor one is not prompting. It is specification, and specification is the same skill that separates good delegation from bad delegation with people. Vague instructions produce work that is technically responsive and not what you meant, and that is true whether you are briefing a contractor or an agent.
What that looks like in practice
Cowork is built for work with several steps and real inputs rather than a single question. Turning a folder of receipts into a formatted expense report. Pulling themes and action items out of a stack of meeting transcripts. Producing a spreadsheet with working formulas rather than a CSV you then have to fix. Building a deck from rough notes.
Two things about it change how you work rather than only how fast.
Sessions continue without you. Work runs in the cloud, so it keeps going when you close the laptop, and you can open the same session on your phone. For anyone whose day arrives in twenty-minute pieces rather than afternoons, that is the meaningful part.
Tasks can be scheduled. You can set work to run on a cadence rather than starting it each time, which is the difference between a tool you use and a system that runs.
When to use which
| Chat | Cowork | |
|---|---|---|
| Best for | Thinking, drafting, deciding | Multi-step execution |
| You are | In every turn | Reviewing at the end |
| Clarity needed up front | Low, work it out as you go | High, specify first |
| Typical use | A question, an argument, a draft | A task across several files |
| Usage cost | Standard | Higher, because more is happening |
The usage difference is worth knowing before you start. Cowork consumes more of your allocation than chat, because multi-step work is doing more. Using it for something a single question would answer is an expensive habit to build.
What it does not do yet
What Claude remembers about you in chat does not carry into a Cowork session, and within Cowork, memory works inside projects only. So the context you have built up in chat is not automatically present, and for now you either work inside a project or supply the context with the task.
That constraint will move. The general point behind it will not, which is that the quality of what comes back is set by the quality of what you specified, and there is no version of this where vagueness produces good work.
If you want the specification skill rather than the tool, start with what to automate with AI first, because a context brain is what makes any of this produce something that sounds like you.
A note on what Cowork is not
Cowork is a general knowledge-work tool. It is not a coding platform and it is not tied to any particular build environment. If what you are doing is building and deploying software, that is a different question, and we have covered it in what Replit is and how it is different from Claude.