What Claude Cowork is,
without the marketing.
Anthropic’s own product page contains a line worth reading twice: Cowork is not a chat assistant. Most of the confusion about it clears up the moment you take that literally.
Claude Cowork is described by Anthropic as a system that executes multi-step knowledge work on your behalf. It runs on the desktop, in local files and folders, and it is available on all paid Claude plans. Short description, long implications, because a tool that returns finished work needs a different kind of instruction from one that returns an answer.
It works where your files already live
Claude Cowork runs on the desktop, in the local files, folders and applications you use every day, rather than in a browser tab you paste things into. That one design decision is why it can hand back a finished deliverable. It reads the files in the folder you point it at and writes the result back as a file you can open, which removes the step that quietly eats the afternoon. Nothing is pasted anywhere, and nothing is reformatted twice.
You brief an outcome, not a message
The unit of work has changed. A chat prompt asks for a response that you then do something with. A Claude Cowork task describes a finished state and lets it work out the steps. So the hard part of the brief is defining done: what the output is, what format it takes, and which cases it should flag for you rather than decide on its own.
Your folder structure is now part of the prompt
The preparation that pays is not clever phrasing, it is file hygiene. Names that mean something, one folder per job, and no seven versions of final. Every ambiguity in the folder becomes an ambiguity in the work. You can usually predict how a run will go by looking at the file names before it starts.
It is built to stop at the decisions
Anthropic is explicit that consequential decisions remain with the user and that the product is designed with human oversight in mind. In practice, the review step is yours and it does not go away. Read what it produced rather than the summary of what it produced, because the summary is written by the same system that did the work. On a first run, spot-check the cases you would have found hardest to decide yourself, since those are where the judgement calls are hiding.
Who gets the most out of it
The target is work that is time-consuming but not technically complex: organising files, preparing documents, synthesising research, pulling data out of unstructured sources. Analysts, researchers and legal teams tend to recognise their own week in that list. Developers already have Claude Code. Claude Cowork is for people whose deliverable is a document rather than a diff.
If you find yourself watching each step and correcting it mid-flight, you have written a chat prompt and pressed it into a task shape. Stop and write down what the finished thing looks like instead: the file, the format, and the rule for anything ambiguous. Claude Cowork training is mostly supervised practice at that one skill, against your own workload.
Things people
usually ask.
No. It is aimed at people whose work is time-consuming rather than technically complex, and it runs through the desktop app rather than a terminal. The skill it asks for is describing a finished outcome precisely, which is a briefing skill rather than a coding one.
It is available on all paid plans through the Claude desktop app. Usage counts against the same limits as the rest of your Claude use, so a long unattended task competes with your chat and your coding inside the same window.
It works on the files and folders you give it access to, and Anthropic designs it so that consequential decisions stay with you. Treat folder access the way you would treat handing a new contractor a shared drive: scope it, review the output, and check your own organisation’s policy first.