If you have used Claude in a chat window, Cowork will feel familiar for about five minutes and then stop feeling familiar. The difference is not that it is smarter. The difference is that it has somewhere to put things.
What Cowork actually changes
In a chat, everything lives in the conversation. You paste text in, you copy text out, and when the conversation ends it is gone. Cowork gives Claude a workspace: it can open your files, work across several of them, produce new ones, and keep going on a task while you do something else.
That changes what is worth asking for. In chat, the natural request is "write me a paragraph". In Cowork, the natural request is "go through these eleven files and build me the summary table". The second one was always technically possible in chat, it was just so painful to feed in the material that nobody did it.
What it is genuinely good at
Three shapes of work, based on what people actually come back to it for.
Reading more than you want to read
A folder of contracts, a quarter of invoices, forty survey responses, a long report you need three facts out of. Cowork will work through the whole set and produce something structured. This is the single most useful thing it does, and it is the thing chat was worst at.
Producing a file rather than a message
A report, a deck, a reconciliation, a spreadsheet with the analysis in it. The output is a document you can send to someone, not a block of text you have to reformat first.
Work that takes longer than your attention span
Research that needs a dozen sources checked, a data set that needs cleaning before it means anything. You can start it and come back.
The four mistakes beginners make
1. Treating it like a faster chat
Asking one-line questions gets you one-line answers, and you have gained nothing over chat. The value appears when the task involves several files or several steps.
2. Not saying what the output should be
"Analyse this" produces an essay. "Produce a spreadsheet with one row per invoice, columns for supplier, date, net, VAT and a flag for anything missing a PO number" produces the thing you wanted. Name the artefact.
3. Assuming it knows your conventions
It does not know that your fiscal year starts in April, that anything over 5,000 needs two approvals, or that the client hates the passive voice. None of that is inferable. It has to be told, and if you are going to tell it every time, that is exactly what a skill file is for.
4. Not checking the numbers
Treat arithmetic as something to verify, always, against your own system. Cowork is reliably good at structure, explanation and catching what nobody asked about. Reconciling to the penny is not the same skill.
Where skills come in
A skill is a file of standing instructions: what role to work in, what to ask before starting, what format to produce, and what never to guess at. You enable it once and it applies to every session, so you stop re-explaining yourself.
One detail that catches people out, and it is worth knowing early: Cowork does not read the ~/.claude/skills/ folder on your computer. It loads the skills enabled on your Claude account, which sync when a session starts. So enabling a skill once in your settings is enough, and hunting for a folder to drop a file into is a dead end. The full explanation of how skills work covers this properly.
A sensible first week
- Day one: give it a real folder of your own files and ask for a summary table. Not a test task, a real one. You will learn more from one real task than from ten toy ones.
- Day two: ask for an output file rather than a message. A report, a deck, a spreadsheet. Notice how much of the formatting work disappears.
- Day three: notice what you had to explain twice, and write it down. That list is your first skill.
- End of week one: enable a skill for whatever you do most, and stop re-explaining it.
Interactive sessions and scheduled sessions
Cowork runs in two modes, and the difference matters more than it first looks.
An interactive session is the one you are sitting in front of. You can answer a question halfway through, correct a wrong assumption, or stop the run when it heads somewhere you did not intend.
A scheduled session runs without you. It starts on a schedule, does the work, and leaves the result behind. Nobody is there to answer anything.
That second mode is where most people get caught out, and the reason is skills. A skill that opens by asking "which quarter are we reporting on?" is helpful when you are there to answer. In a scheduled run the same question has nobody to answer it, and the run either stalls or guesses.
The practical rule: if you intend to run something on a schedule, the skill needs a sensible default for every question it would otherwise ask, and it should state in the output what it assumed. Then a scheduled run produces a file you can check, rather than a file you have to distrust.
Skills themselves load identically in both modes. They come from the skills enabled on your Claude account and sync when the session starts, so you enable a skill once and both modes see it.
Where to start
If you would rather start from something already tested than write your own, we publish skills built around file and document work:
Claude Cowork skills →
Free skills, to test the workflow first →
How to install a skill, step by step →
What a skill is and how to enable one →
KissMySkills is an independent marketplace and is not affiliated with, endorsed by, or connected to Anthropic, OpenAI, or any other AI company. Claude and Claude Cowork are products of Anthropic.