The short answer
Claude Cowork is where you hand over a whole job rather than ask a question. It automates multi-step workflows that pull from your own files, folders and applications, and it keeps going without you standing over it prompting each stage. It launched in January 2026, built primarily for people who do not write code, and Anthropic described it at the time as Claude Code without the code.
The feature to build your understanding around is the one at the end: Claude asks for your final review and approval before shipping the work. The tedious middle happens without you. The consequential last step does not.
Cowork is young and moving fast. It went to the cloud in July 2026, the cloud beta started with Max subscribers, and Anthropic is unifying Claude chat and Cowork into a single space with access to the same files. Anything written down about availability, this article included, will drift. Check Anthropic's own pages before making a decision that depends on it.
What is the difference between answering and doing?
Most people never make this switch with AI. They keep asking for advice about the work instead of asking for the work.
In a chat window, you are the engine. You paste something in, it responds, you read it, you fetch the next thing, you paste that in. Nothing moves unless you move it, and if you get distracted for an hour the job sits exactly where you left it. Useful, but you are still doing all the walking.
Cowork inverts the arrangement. You describe the job once, including what finished looks like, and the work continues under its own steam across several connected steps, where later steps depend on what earlier ones turned up. You come back to a result rather than a conversation.
The difference sounds academic until you put a real job in it. Asking how you should structure a monthly client update gets you a sensible essay about structuring monthly client updates, which you then have to go and act on. Asking for the monthly client update itself, assembled from the folder where the evidence actually lives, gets you a draft with your project names in it. Same tool underneath. Completely different afternoon.
In plain English
- Cowork:
- Claude's environment for handing over a whole job rather than a single question.
- Multi-step workflow:
- A job made of several connected actions, where later steps depend on what earlier ones found.
- Unattended run:
- Work that carries on while you are not watching it.
- Approval step:
- The moment Claude stops, shows you what it produced, and waits for your yes before anything ships.
- Scheduled task:
- A job set to run at a chosen time, which can happen while your own device is off.
What does it do with your files?
This is the part that makes it feel different from a chat window, and it is also the part that deserves a moment of thought before you switch it on.
Cowork works from your own material. Your files, your folders, the applications you already use. Which means the answer to "where did this figure come from" is your actual tracker rather than something plausible-sounding, and the answer to "did you read all twelve documents" is either yes or a named list of the ones it could not open.
That closeness to your real work is the value. It is also why the first job you hand over should be one that only reads. Nothing created, nothing sent, nothing renamed, nothing deleted. You learn how the thing behaves on a job where the worst possible outcome is a document you delete afterwards.
It is worth being deliberate about which folder you point it at, too. Most people's working directories are an archaeological site: the current version, the version before the client changed their mind, a copy someone renamed FINAL in 2024, and three things that should never have been saved there at all. A person opening that folder applies years of context about which file is real. Something reading it fresh applies whatever you told it to apply, which is an excellent argument for telling it plainly which source counts and what to do when two of them disagree.
Why the approval step is the whole product
Claude asks for your final review and approval before shipping the work. It sounds like a small piece of product design. It is the entire safety model.
Think about what it means structurally. The middle of a job, which is where the tedium lives, happens without you. The end of a job, which is where the consequences live, does not happen without you. Anything that leaves a mark on a world other people can see, sending, publishing, filing, telling someone something, waits at a gate you control.
The catch is that an approval step is only as good as your ability to actually approve. A gate you wave everything through is not a safety feature, it is a formality with a nice interface. Which brings us to the part of the brief you write yourself.
โ Weak prompt
Prompt
Go through my receipts folder and sort out the expenses summary.
Output
A tidy summary appears. You do not know which files it read, what it did with the two receipts that had no date, or whether it converted that currency using today's rate or last month's.
Sort out is not a finished state. Nothing here says what done looks like, so when the work arrives you have no way to check it and the approval step becomes a rubber stamp.
โ Good prompt
Prompt
Read every file in my receipts folder for June. Produce one table: supplier, date, amount, currency, category. Take every figure exactly as printed and never convert a currency. If a receipt is unreadable or missing a date, list it separately as NEEDS ME rather than guessing. Do not rename, move or delete anything. Show me what you read, what you could not read and anything you assumed, then wait.
Output
A table of forty-one receipts, plus three listed under NEEDS ME with the reason, plus one noted assumption about a supplier name that appears two different ways.
Now approving means something. You know the sources, the gaps are named rather than smoothed over, and the decision to accept it is genuinely yours.
The pattern to steal is the last line. Ask for what it did, what it could not find, and what it assumed. The third one is the one people forget and the one that matters, because assumptions are where a helpful job and a quietly wrong job part company.
What should you hand over, and what should you keep?
GOOD CANDIDATES
reading a pile of things and producing one summary
the same shaped job that comes round every week
cross-checking two sources against each other
work where every line can be verified in under a minute
KEEP THESE YOURSELF
anything going to a client or the public without your eyes on it
judgement calls you would struggle to explain afterwards
jobs where you cannot tell a good result from a plausible one
anything you could not undo by Monday morning
The bottom line of that second column is the useful one. Not how clever is it, not which model is underneath. Just: after this has finished, what is different, and can I put it back?
The hard part nobody mentions
Since July 2026 Cowork runs in the cloud. A task can start on your laptop, carry on in the background and be reviewed from your phone hours later, even after you closed the app. Scheduled tasks can run while your device is off entirely.
The convenience is obvious. The consequence is worth sitting with, because it is the real skill this product asks of you.
You are now approving work whose middle you did not watch. Possibly on a small screen. Possibly between two meetings. Possibly seven hours after the thing that produced it made a small decision at step two that turned into a confident conclusion by step nine. Nobody was in the room for the interesting part, which is precisely how something arrives well organised, fluently written, correctly formatted and wrong in a way that is now buried three layers down.
The approval step survives the move to the cloud perfectly well. Your ability to approve carefully is the thing under pressure. In practice that means insisting on jobs whose output you can spot-check quickly, asking for the assumptions list every single time, and being honest with yourself about which approvals you are actually reading and which you are just waving through while walking to the lift.
Start small, start read-only, and widen the boundary once you have seen what it does with a narrow one. Our free Claude Cowork course covers what to hand over, how to brief a job so the result is checkable, and how to review work you did not watch happen. If you want the broader habit of delegating whole jobs rather than asking questions, AI That Does the Work is the companion, and context engineering is where briefing properly gets taken apart. Not sure where to begin? Start here.