Every Claude Cowork Concept Explained

Projects

A walled-off workspace for one body of work. A project holds its own files, its own instructions and its own memory. The problem it solves is one job’s settings turning up in another job: the way you like invoices laid out for a building firm arriving in the middle of the report you are writing for a dentist.

In plain wordsMemory works the same way inside a project. Each project keeps a memory that belongs to it alone, so the more you use one project the sharper Claude gets inside its walls and the less it carries anywhere else.

In practiceYou make one by picking a folder, naming the project, choosing where on your computer it is stored, and adding the instructions that belong to it. After that you open the project rather than going hunting for the folder. The person who feels the difference is anybody running more than about three streams of work at once: one client, one internal thing and one side project is the point where keeping it all in one place starts costing you time.

Watch outThree limits worth knowing before you build anything on them. Projects live on your desktop and are stored on your own machine. There is no cloud copy of them. And on the team plans they cannot be shared with anybody else, so if you wanted to set a colleague up with the same working arrangement you have, a project is not how you hand it over today, and you would be writing the instructions out for them by hand.

Skills

A set of written instructions for one job you do again and again, saved so that Claude can pick it up and follow it. A skill sounds like programming and it is not: a skill is a folder with one text file in it, and that file is written in ordinary sentences, the way you would brief somebody joining your team on Monday.

In plain wordsA skill takes the place of the checklist you wrote once and stopped opening. The difference between a skill and that document is that a skill gets carried out rather than remembered.

In practiceWhat a skill suits is the task with steps you always forget in the same order. Checking a deck before it goes out, formatting the monthly report, running the same five checks over a spreadsheet before anybody else sees it. Written out, a skill for checking a deck says what to look at, in what order, and what counts as a failure. Every slide has a heading. No slide has more than six lines. Every figure has a date beside it. The client’s name is spelled the way they spell it.

WhereYou turn skills on under Customize in the sidebar, and the ones you write yourself stay private to your own account, unless you are on a team plan and choose to share them.

Watch outThe common mistake is building one enormous skill that does everything. Anthropic’s own guidance says the opposite, which is separate skills for separate purposes, because a skill trying to cover four jobs gets picked for the wrong one.

Slash commands

The forward slash key followed by the name of a skill, and typing it runs that skill on the spot. Claude will pick a skill up on its own when your request matches it; a slash command is how you call one by name.

In plain wordsOne slash command comes built in already, and it is the one most people use first. Typing slash schedule inside a task sets that task to run again on its own.

In practiceThe reason to name a skill rather than describe the job is certainty. Typing the name means you know which set of instructions is about to run, instead of hoping the words you chose matched the right one. It also saves you having to remember your own phrasing: without it you are recalling roughly how you asked last time and hoping it lands on the same instructions, and with it you press one key and read the list. Which makes what you call your skills matter more than it sounds. A skill called check is useless in a list of nine. A skill called check client deck before sending tells you what it does a year later, when you have forgotten writing it.

Watch outThe time not to reach for the slash is when you want Claude choosing. If you are not sure which of your skills fits, describing the job in your own words lets Claude match it, and that is the better move on anything you have not done before.

Plugins

A bundle. A plugin packs skills, connectors and sub-agents together into one thing that installs in a single action. What it gives you is somebody else’s finished working setup, rather than their instructions for building one: you install the bundle and the whole arrangement is already wired together.

In plain wordsTwo of those words have their own entries further down this sheet. A connector is a link between Claude and another program you already use. A sub-agent is a second copy of Claude working alongside the first one.

In practiceThe day this changes is the setup day. Somebody joins, and instead of a morning spent installing pieces and copying instructions into the right boxes, they install one bundle and start on the work.

WhereYou browse plugins in the same directory as skills and connectors, under Customize.

Watch outA plugin is something somebody else wrote, so the same caution applies as installing anything else. It can carry connections to your own programs and instructions you have not read, so find out where it came from before it is running inside your work. And for one person doing one job on one machine, a plugin is more than the situation needs: a skill you wrote yourself does that job, and the bundle earns its place when the arrangement has to travel to somebody else.

Connectors

A link between Claude and another program you already use, so that Claude can read from that program and do things in it. Up to this point everything has happened inside one folder on one computer. A connector is what takes Claude outside that boundary. Each one is switched on by you, one at a time, and none of them is on until you switch it on.

In plain wordsYou will see these described as M C P connectors. It stands for the Model Context Protocol, an agreed way for a program to describe what it can do, so that any assistant can use it without somebody writing a special connection for that one pairing. You never touch the standard yourself. It is the reason the list of available programs keeps growing.

In practiceWhat it changes is the work that used to be assembly by hand. Take one Monday morning with a few connected: Claude reads the weekend’s messages in the shared channel, opens the sales sheet, checks what is in your calendar for the week, and writes you the one page you were going to spend an hour building before your first meeting, without you carrying anything between the three yourself.

Watch outThe thing to think about before switching one on is what it opens up. Connecting your email gives Claude your email, so the sensible order is to connect the one program the job needs, rather than everything you own on the first afternoon.

Chrome extension (Claude in Chrome)

An add-on for the Chrome web browser that puts Claude in a panel beside the page, where Claude can read what is on that page and click through it for you. A connector works with a program that offers a connection, and most of the internet does not offer one. This is what covers the rest.

In plain wordsThere is a second way to get Claude onto a web page, and it helps to know which is which. Cowork now has a browser built into it, so a task can open a site without you installing anything at all. The Chrome add-on is the one you reach for when you want Claude working in the browser you are already sitting in.

In practiceIt covers the sites that never built a connector, which is most of them. A supplier portal, a booking system, an internal tool with one login that nobody has updated since it was made. Inside those pages Claude fills forms, moves through steps and finishes tasks. Claude can take actions without asking permission for every single click, and each action is checked before it happens against what you asked for. The job this takes off somebody is the twenty-minute portal: log in every week, work through six screens, copy the same figures out.

WhereIf you looked for Claude in Chrome before and could not get it, there is a reason. As of the film it had just become available on every paid plan.

Watch outWhat limits the add-on is where it runs. It does not work in other browsers built on the same underlying technology as Chrome, and it does not work on phones.

Computer use

Claude working your actual desktop. Claude looks at the screen, moves the pointer, clicks and types into the programs you have open, the way a person sitting at your machine would. A browser is one program on your screen, and the spreadsheet, the accounts package and the design tool sitting open next to it cannot be reached that way.

In plain wordsComputer use is a research preview, which is Anthropic’s own label for something released early and still being worked on. It runs on Mac and Windows, on the Pro and Max plans.

In practiceThe reason it exists is the software that has no other way in. The old desktop program with no connector, no web version and no way to reach it except the buttons on the screen, which describes a great deal of what businesses actually run on. Worked through once: the stock system is a program from two thousand and nine that runs on your machine and nowhere else. Claude opens it, types the reference into the search box, reads the level off the screen, and writes that number into the spreadsheet in your folder. Practice management software, a warehouse system, the accounts package the business has run since before anybody thought about connections.

Watch outClaude asks before touching each application, and some categories are shut off from the start, including investment and trading platforms and anything to do with cryptocurrency. It also differs from everything before it in one way that matters: when Claude runs code, it runs inside that sealed container. When Claude uses your computer, there is nothing between Claude and what is on your screen. Which makes the sensible guidance narrow. Computer use is for the task you cannot do any other way, on a machine where you are content to have Claude clicking, and it is not the tool to reach for when a connector would have done the job.

Scheduled tasks

A job you set once and it runs on its own after that, daily, weekly or monthly, at the time you picked.

In plain wordsThe part that surprises people is where these run. Scheduled tasks run in the cloud, on Anthropic’s servers, so your computer does not have to be awake and the Claude app does not have to be open.

In practiceA scheduled task also gets everything a normal task gets, which means the programs you connected, the skills you turned on and the plugins you installed are all there when the task fires at six in the morning. The one worth setting up first is the report you build by hand every week: Claude reads the week’s numbers, writes the summary in the shape you always write it, and it is finished before you are at your desk on Monday.

WhereYou set one with the slash schedule command, or from the scheduled list in the sidebar, and that same list is where you go to change one or switch it off.

Watch outWhere this goes wrong is the task that keeps running after the reason for it has gone. A weekly report on a project that finished in June is still arriving in September, so a look through that list every so often is the fix. Anthropic’s own safety guidance carries a caution too, and it follows from the thing that makes them useful: a task running when nobody is watching is a task nobody is watching, so the ones you leave unattended should be the ones where a mistake costs you nothing.

Dispatch

Dispatch lets you send Claude a job from your phone and have it done on your desktop computer at home, using the files and the programs that are on that machine. Setting it up is a pairing between the two devices, and after that it behaves as one conversation: you send from the phone, you watch Claude work from either device, and you pick it up on the desktop later in the same thread.

In plain wordsThe difference between this and a scheduled task is the thing to carry out of both. Scheduled work runs on Anthropic’s machines, so the lid can be shut. Dispatched work runs on yours, so it cannot.

In practiceThe moment Dispatch is built for is the one on the train. You think of the thing that needs doing, your computer is at home with all your files on it, and the work starts now instead of at nine tomorrow when you are in front of it.

Watch outOne hard requirement: your computer has to be awake with the Claude desktop app open, because the work happens on that machine and nowhere else. It runs on the Pro and Max plans, it is called Dispatch rather than dispatch mode, and it holds one continuous thread, so there is no second conversation to start while the first one is going. And if your desktop runs Linux, which is a third operating system used alongside Windows and Mac, the jobs that need Claude to work your screen are not available over Dispatch.

Sub-agents

A copy of Claude working on one part of your job at the same time as other copies work on the others. Claude decides to do this on its own when a job is big enough to divide: you hand over one brief, and underneath it the work is being cut up and shared out without you arranging any of it.

In plain wordsEach of those copies gets its own context window, the bounded working surface from concept five. That is the whole reason this is faster, because four separate surfaces hold four times as much as one.

In practiceThe shape of job this suits is the one whose parts do not depend on each other. Researching six competitors is six jobs that never need to talk. Writing a document where each section follows from the one before it is one job, and cutting that up makes it worse. What comes back to you at the end is one finished piece of work, not four: the pieces are brought together before you ever see them.

Watch outThat speed has a cost, and it is a plain one. Running several copies at once uses your allowance several times over, because each copy is doing real work of its own.