Answering a question is not doing the job
There is a switch most people never make with AI. They keep asking for advice about the work instead of asking for the work itself. Our course on delegating pulls that habit apart properly and it is worth your time. This lesson is about what happens when a product is built around that idea from the start rather than left as something you have to remember.
Claude Cowork launched in January 2026. Anthropic built it primarily for people who do not write code, and the description used at launch was blunt: Claude Code without the code. That phrase does more work than it looks. The name it borrows from belongs to Anthropic's tool for developers, and what the comparison points at is a shape rather than a feature list: work that runs across many connected steps and comes back finished, instead of a conversation you have to keep feeding. Cowork puts that shape within reach of ordinary office work.
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.
What it actually does
Cowork automates multi-step workflows using your own files, folders and applications, without you standing over it prompting each stage. That last part is the whole difference. In a chat window you are the engine. You paste, it responds, you read, you paste again, and nothing moves unless you move it. In Cowork you describe the job once and the work continues under its own steam.
This product is moving quickly. It went to the cloud in July 2026, the cloud beta started with Max subscribers and is expanding to more plans, and Anthropic is unifying Claude chat and Cowork into one space with access to the same files. Anything written down about what is available on which plan, including anything on this site, will drift out of date. Check Anthropic's own pages before making a decision that depends on availability.
The bit that defines it
Here is the feature to build your whole understanding around. Claude asks for your final review and approval before shipping the work.
That sounds like a small piece of product design. It is the entire safety model. Cowork is deliberately built so that the useful, tedious middle of a job happens without you, and the consequential end of it does not happen without you. Everything else in this course hangs off that one sentence.
โ Weak prompt
Prompt
Have a look through my project folder and sort out the client update.
Output
A document appears. You do not know which files it read, what it assumed about the two conflicting version numbers, or whether the figures came from the current tracker or last quarter's.
Sort out is not a finished state. Nothing here says what done looks like, so the approval step becomes a rubber stamp on work you have no way to check.
โ Good prompt
Prompt
Produce the monthly client update from the files in my Northgate folder. Three sections: delivered, in progress, blocked. Every date and figure taken from the tracker only, quoted exactly as written. If a status is missing, write NOT FOUND rather than estimating. Do not send it anywhere. Show me what you read, what you could not find and anything you assumed, then wait for me.
Output
A draft with three named sections, plus a short list of the two items where the tracker had no status and one assumption about a date format.
Now the approval step means something. You know the sources, you can see the gaps, and the decision to ship is genuinely yours to make.
Checkpoint
Cowork does the middle of the job without you and stops before the end. The approval step is not a formality bolted on afterwards. It is the design.
Why the cloud move matters to you
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 later, even after you have closed the app. Scheduled tasks can run while your device is offline entirely.
The convenience is obvious. The consequence is the part worth sitting with. You are now approving work whose middle you did not see, possibly on a small screen, possibly hours later, possibly while walking between meetings. The approval step survives the move to the cloud perfectly well. Your ability to approve carefully is the thing under pressure, and lesson five exists to protect it.
A QUESTION A JOB
one exchange several connected steps
you supply the material it goes and finds the material
you assemble the result it assembles the result
you are present throughout you are absent for the middle
nothing to approve approval is the last step
Read the documents in [one folder you choose].
Produce a single page listing, for each document:
what it is, who it is from, and the most recent date inside it.
Take dates exactly as written. Never estimate one.
If a document is unreadable or empty, say so on its own line.
Do not create, move, rename or delete anything.
Show me the result and wait.
Before you finish, show me three things:
WHAT YOU DID the steps you actually took, in order
WHAT YOU COULD NOT FIND anything missing, named
WHAT YOU ASSUMED every gap you filled in yourself
Then stop and wait for my approval. Do not send or publish anything.
Start with a job that only reads. Nothing created, nothing sent, nothing changed. You learn how the thing behaves and the worst possible outcome is a document you delete.
๐ Quiz
Question 1 of 4How was Claude Cowork described when it launched in January 2026?