If you want to know how to use Claude Cowork for a small business, the first session takes about a minute. Open the Claude app, pick "Cowork" in the message box instead of "Chat", describe a job, and watch it plan the work and do it.
Starting Cowork is not the hard part. Deciding what it is allowed to touch is.
I run three companies on this setup from a Windows machine in Calgary. Cowork reads my inboxes and calendar, works inside my CRM, drafts my replies, and runs a morning briefing that tells me where the money is. It also, for several weeks this year, silently stopped running a job I was counting on. This is the whole setup, including that failure and what I built so it cannot happen quietly again.
Give Cowork one folder, one connector and a written set of operating rules before the first real task. Turn each repeating job into a plain text skill file, schedule it from the "Scheduled" page, and keep a visible board that shows when each job last produced output. The rules and the visibility matter more than the tool.
Last verified against Anthropic's Cowork documentation on 12 September 2026.
Before I sold anything I was an avionics technician on F-18s in the RCAF. Every job on that aircraft had a written procedure, because the same result had to come out twice, from different hands, on a bad day. That is the problem with handing work to an AI. Cowork is a capable colleague with no memory of your business and no sense of what it must never do. The procedure is what makes it safe to leave alone.
When should you use Claude Cowork instead of chat or Claude Code?
Anthropic ships three ways to use the same model and most owners pick by accident. Here is the split I use, and it matches what the official getting started guide says each one is for.
| Use this | When the job is | What it can touch |
|---|---|---|
| Chat | A question, a rewrite, a quick think. One answer. | What you paste in. Cheapest on your usage. |
| Cowork | A multi-step job with a deliverable: triage this inbox, prep this meeting, run this every morning. | Folders you connect, connectors you approve, skills, plugins, scheduled tasks. Uses more of your allowance than chat. |
| Claude Code | Building something: a connector to an app with no ready-made integration, a dashboard, a script on your own machine. | Your file system and terminal. Add it later without redoing anything. |
I build in Code and run in Cowork. Any owner who is not technical should live in Cowork.
Cowork is on every paid plan: Pro, Max, Team and Enterprise. It runs in the desktop app for Windows and Mac, on the web, and on mobile. But three things only work while the desktop app is open: reading your local folders, connectors that run on your own machine, and letting Claude use your browser or screen. That matters in Step 5.
How to use Claude Cowork for a small business: my real setup
Nothing exotic. One Windows laptop, Claude Desktop with Cowork, and Claude Code for build work. What makes it run a business is the layer of connections and written rules on top.
| Layer | What it is | In my setup |
|---|---|---|
| Connectors | Live links from Claude to tools I already use | One connector per GoHighLevel sub-account (I run five), Gmail, Google Calendar, Google Tasks, my meeting recorder, and the team's task board |
| Operating rules | Global instructions that apply to every session | Drafts only, never send. Mark as read only, never archive or delete. Never move a deal in a rep's pipeline. Timezone is always Mountain. No passwords or keys in chat. |
| Skill files | One plain text file per repeating job | Daily briefing, weekly blog run, sales call report, client onboarding, connector health check |
| Scheduled tasks | Jobs Cowork runs on a cadence | Morning briefing, a weekday rep report, a scan that refreshes my dashboard |
| The cockpit | One HTML page I open every morning | Today's briefing, inbox items needing a decision, the pipeline, and a health board showing every scheduled routine and when it last ran |
Read the second row again. Those five rules are the only reason I can let the thing run unattended, and every one was written after something went slightly wrong. When I do an AI build for an owner, the rules get written before a single connector goes live.
Step 1: Start a Cowork session and give it one folder
You need the latest Claude Desktop app. If it is not installed yet, the full install walkthrough covers every screen. Then:
- Open Claude Desktop and sign in with the account your business uses.
- In the message box, select "Cowork". Chat and Cowork share the box; you are switching modes, not apps.
- Connect a folder. A dedicated working folder, not your whole Documents drive. Anthropic's safety guidance says the same: keep financial documents, credentials and personal records out of it.
- Pick how Claude checks with you, using the mode selector in the chat box. Manual pauses for approval on each action (older builds call it "Ask before acting"). Auto reviews its own actions and carries on. Skip runs with no checks. Start on Manual. Move to Auto for jobs you have watched succeed. I never use Skip on anything that touches a customer.
- Describe one job with an ending. Not "help with marketing". "Read the last twenty emails in this inbox, list the ones that need a reply from me, and write each reply as a draft. Do not send anything."
Cowork has projects, created from the "+" beside "Projects" in the left panel. A project holds its own instructions, context, memory and scheduled tasks. If your business context already sits in a Claude Project, import it and skip most of the setup. One catch: a project built from a local folder stays on that machine and is not saved to your Claude account.
Step 2: Connect the one tool you live in. One.
Connectors are managed from the "+" menu in the chat box or from Customize > Connectors. Every owner wants nine connections on day one. Connect the one system you are in most, use it for a week, then add the next.
For me that was the CRM. Most of what my business does, from a cold-call follow-up to an invoice, lives in GoHighLevel, so one working connection to it moved more than five half-working connections to everything else. Each sub-account gets its own connector, so when one breaks the rest keep running, and a job reading one company's contacts cannot write to another's.
Two rules for any owner. Block anything that moves money or holds health, banking or personal records; Anthropic names those categories directly. And stick to connectors from the Claude directory or ones your own technical person built. A connector you cannot explain is one you should not run.
Step 3: Write the operating rules before the first real task
Go to Settings > Cowork, find Global instructions, click Edit, and write what Claude must never do in any session. Mine are the five in the table above. These four belong in almost every small business:
- Drafts only. Claude writes the email, the owner sends it. Nothing leaves without you.
- Read and summarize, never delete. Cowork already asks before permanently deleting files. Extend that to email, records and contacts in your own words.
- Never change something a person owns. A rep's deal stage, a booked appointment, a colleague's task. Recommend the change, do not make it.
- Secrets never go in chat. Passwords and keys live in the connector setup, never in a message.
Folder instructions are the same idea scoped to one folder. Cowork updates those on its own during a session, which is useful and worth reading occasionally to see what it decided to remember.
Step 4: Turn one repeating job into a skill file
This step separates the owners who get value from the ones with an expensive chat window. A skill file is a plain text markdown file, one per repeating job, written the way you would brief a new hire: what triggers it, what it reads, the steps in order, the hard rules, and what the output must look like. Cowork's plugins bundle ready-made skills with connectors, but the ones that change your week are written around your business.
Write the daily briefing first. It is the job every owner already does badly each morning, across four tabs, in the wrong order. This is the prompt I would give Claude to write yours. It asks questions first, then hands you back a finished file.
ROLE: You are an operations assistant who builds written procedures for small business owners. You write plainly and never invent tools the owner has not named.
CONTEXT: I run a small business and use Claude Cowork with a few connected tools. I want a daily briefing that runs every weekday morning and tells me, in one scan, what needs my attention and where the money is. It must only read, summarize and draft. It must never send, delete, archive or move anything.
TASK: Before writing, ask me up to five questions, one at a time: which tools I have connected (email, calendar, CRM, task list, anything else), what time I read it, my two or three revenue priorities right now, the three kinds of item that must always appear at the top, and what I never want it to touch. Then write my daily briefing procedure from my answers.
FORMAT: Return one complete markdown file in a single code block named daily-briefing.md. It must contain: a name and the trigger phrases I can say to run it; the sources it reads, in order; the mandatory sections of the output, each with a one-line rule for what belongs there; the hard rules it must never break; and a closing line naming the one action I should take first today. Under 400 words. No filler.
Paste that into Claude and it will ask you up to five questions before it writes anything. Answer them properly, in your own words. Then it hands you back a finished file built around your business and the tools you actually have connected, not around our example.
That file is the thing worth keeping. Save it as daily-briefing.md. The .md on the end just means markdown, which is a plain text file with a little structure, and it is the format AI tools read best. If that is new, we cover it in what an MD file actually is.
Now open a Claude Project and add that file under project instructions, or upload it as project knowledge. Every chat you start in that project begins already knowing it. You stop re-explaining yourself every morning, which is the entire point.
Once that file exists, every other job follows the same shape. Written once, run by asking, or run on a schedule.
Step 5: Schedule it, then make it visible
Click "Scheduled" in the left sidebar, then "New task". Anthropic's scheduled tasks guide documents two paths: "Create with Claude", where it interviews you and you click "Schedule" at the end, or "Set up manually", where you enter the name, prompt, approval mode and frequency and click "Save". Frequencies are hourly, daily, weekly, on weekdays, or manually. The same page shows past runs and lets you edit, pause, delete or run a task on demand.
Two facts the setup screens do not shout about. Scheduled tasks run remotely on Anthropic's side, so they fire whether or not your laptop is awake. And because they run remotely, a scheduled task cannot be tied to a folder on your computer. Anything the job needs must reach it through a connector, a skill or plugin, or a Cowork project. If your briefing procedure lives only in a folder on your desk, the scheduled version cannot see it.
The task that went quiet for weeks
Earlier this year I had a morning routine scheduled in Cowork on Windows. For a stretch it ran. Then it did not, and nothing told me. I only noticed when I went looking for output that should have been there and found the last good run was weeks old. Background runs on Windows were unreliable on my machine at the time, and a virtual machine fix solved one layer of it and not the rest.
The lesson was not "scheduling is broken". Anthropic has since moved scheduled runs to the cloud and the Windows app matches the Mac one. The lesson was that a scheduled job with no visible heartbeat is a liability, because its failure mode is silence. So I changed three things:
- Every scheduled job also runs by asking. The skill file is the job; the schedule is a convenience. If the 7am run is missing, "run my daily briefing" produces it in ninety seconds.
- A health board on the page I already open. My cockpit lists every scheduled routine and colours it by when it last produced output. Green ran today, amber is late, red has been dead for days. A broken robot is visible on the first screen I look at, not buried in a settings page.
- Review every scheduled run, briefly. That is Anthropic's own recommendation and it doubles as the heartbeat. If there is nothing to review, that is the signal.
A scheduled job nobody checks is not automation. It is a delayed surprise.
Step 6: The morning routine, in order
About fifteen minutes, replacing the hour I used to spend across four tabs deciding what to do.
- Open the cockpit. One page: the briefing, the inbox items needing a decision, the deals that have not moved, the health board.
- Read the briefing, or run it. If the scheduled run fired, it is there. If not, I ask for it and it pulls the calendar, the inboxes, the task lists and yesterday's meeting recaps into one scan.
- Clear the drafts. Every email that needed a reply has one waiting. I read, fix, send.
- Take the one money move. The briefing ends by naming a single action that pushes revenue today. I do that one first.
- Glance at the health board. Anything red gets fixed or rerun before I move on.
Connectors so it can see, rules so it cannot hurt anything, skill files so the work repeats, schedules so it happens without me, and a board so I know when it did not.
When something goes wrong
It says "Setting up Claude's workspace" and sits there
Expected. Cowork is updating itself before it starts. Give it a minute. If it is still spinning after several, close the app fully and reopen.
Claude stopped working on my task
A task that uses local folders, a connector on your machine or your browser needs the desktop app open the whole time. Cloud sessions keep going on their own; reopen the session to see progress.
Files are not appearing where I expected
Check that the folder is connected with the right permissions, and check where Claude said it saved the output. It names the location in the session.
My scheduled task did not run, and nothing told me
Open the "Scheduled" page and look at past runs. If the run is missing, run it on demand from that page. If it ran but produced nothing useful, it probably depends on a local folder that scheduled runs cannot reach; move what it needs into a project, a connector or a skill. Then build the heartbeat from Step 5 so you find out the same morning next time.
Cowork or connectors are missing from my app
Almost always plan or permissions. Cowork needs a paid plan and the latest app. On Team and Enterprise plans an admin has to enable it, and internet and browser access sit under Settings > Capabilities.
Frequently asked questions
How do I use Cowork in Claude?
In the Claude Desktop app or on the web, select "Cowork" in the message box instead of "Chat", connect a folder or a tool, and describe a job with a clear end result. Claude plans the work, runs the steps, and checks with you as often as your chosen mode tells it to.
When should you use Claude Cowork?
When the job has more than one step and ends in something: a drafted email, a finished document, an updated record. For a single question or a rewrite, chat is faster and cheaper on your allowance.
Can I use Claude Cowork for free?
No. Cowork is on the paid plans: Pro, Max, Team and Enterprise. Chat works on the free tier; Cowork does not. It also uses more of your allowance per task than chat does, which you can watch in Settings > Usage.
Does Claude Cowork work on Windows?
Yes. It needs the latest Claude Desktop for Windows and has the same features as the Mac app: local folders, connectors, plugins and scheduled tasks. My whole setup runs on Windows.
Can Claude Cowork run tasks while my computer is off?
Scheduled tasks and cloud sessions run on Anthropic's servers, so yes, as long as the task does not need your local folders, a connector on your machine, or your browser. Those three need the desktop app open.
What to do this week
- Monday: Start one Cowork session on Manual with one dedicated folder. Give it one job with an ending.
- Wednesday: Write your global instructions, four rules including drafts only. Then connect the one tool you live in.
- Friday: Paste the briefing prompt above, answer its questions, save the file it hands you, and schedule it for Monday morning. Decide where you will look each day to confirm it ran.
The install takes a morning. The rules and skill files take a week of small corrections. What makes it safe is not the model. It is the written procedure underneath it.
Want this built around your business instead of mine?
ATP AI, Sales and Marketing builds the connectors, the operating rules and the skill files with you, on your machine, then hands you a setup you can run yourself. Calgary based, working with owners across Canada, the United States, Mexico, Panama and the Philippines. Call +1 587-330-3813 or send us a note.
Want more guides like this in your Google results?