In an earlier article, I gave you a complete tour of the Claude ecosystem.
But one section deserved much more than a paragraph: Claude Cowork
It's probably the most underestimated feature of Claude, and yet it's the one that changes how we work day to day the most.
If you only use Claude in chat mode, you're missing the essentials. The chat is an assistant that answers your questions. You copy the text, and you go do the work yourself.
Claude Code is more for developers, more complex to get to grips with.
Cowork is something else: an employee that carries out complete tasks on your computer. You describe the expected result, you go off and do something else, and you come back to a finished file in your folder. Real Excel files with formulas, PowerPoint presentations with clean formatting, Word documents ready to send. Not text in a chat that you have to copy-paste and reformat.
To tell the 3 modes apart, take a look at this cheat sheet:
This article is a complete playbook. We'll configure the space together, understand each building block, and you'll leave with a working setup.
Setting up your Cowork space
Before starting, here's how to access Cowork:
1. Download the Claude desktop app. Cowork is also in beta on the web and on mobile (checked in October 2026), but only the desktop app can open your local folders.
2. Have a paid plan, starting with Pro at $20 a month
3. Open the app and click the Cowork tab between Chat and Code
4. Select the most capable model on your plan and enable extended thinking for better results
October 2026 update: in September 2026, Anthropic started merging Cowork and chat into a single Claude app, on the Pro and Max plans first. If the Chat and Cowork options have disappeared from your message box, you already have the new version. Everything Cowork does is then available from any conversation, with your projects, connectors and skills intact.
Now, the most important part: structuring your working folder.
Forget prompts, think context!
Working with AI models, we've mostly written a lot of increasingly long prompts. More details, more constraints, more examples. And it works, in a classic chat.
Cowork completely changes the logic. The game is no longer the prompt. It's the context you make available to Claude, in the form of files.
The principle: you select a folder on your computer, and Claude can read everything in it. Your documents, your notes, your past examples, your templates. It understands your environment without you having to re-explain everything you've already documented. When those notes already live in a local vault, I detailed the setup with Obsidian. To query a body of sources without building that folder, our NotebookLM tutorial gives the alternative.
Concretely, that means your prompt can be 2 lines. If the folder is well structured, Claude already knows who you are, how you work, and what level of quality you expect.
The whole challenge is then to organize your workspace well, and that's exactly what we're going to see.
It's the most important step. If you remember just one thing: spend 20 minutes on your context files, and every Cowork session will be better.
The folder logic
First rule: never point Claude to all your documents. Cowork has real read and write access to the folder you share. If something goes wrong, you want the damage contained within a limited scope. And above all, a folder that's too broad = too much noise. Claude will read useless files and consume tokens for nothing.
So create a dedicated folder, for example “Cowork”, and give access only to that one.
Inside, it's best to create 3 subfolders:
- CONTEXTE: what Claude needs to know about you before each task. It's the most important folder, the one that turns a generic AI model into a more personalized tool.
- PROJETS: your ongoing projects. One subfolder per project with the brief, the drafts, the reference documents for that specific work. (see example below)
- RÉSULTATS: where Claude drops its finished deliverables. Claude never goes there on its own (it saves tokens), unless you point it to a specific file.
In the CONTEXTE folder, add 3 markdown files (.md):
- a-propos.md: your role, your company, your audience, your current priorities. It's your professional identity card for Claude. No need for a novel: 1 to 2 condensed pages are enough. If the file is too long, Claude summarizes it instead of reading it in detail, and you lose precision.
- tone_of_voice.md: your tone, your style, the words you use, the ones you hate. The writing patterns that make you cringe. If you don't want Claude to write “like an AI”, this is where it happens.
- working-preferences.md: how you like documents to be formatted, how Claude should handle ambiguities, your non-negotiable rules. For example: “always start with an executive summary”, “never more than 15 pages”, “output in .docx”.
In the end, it looks like this:
Important point about document size: keep these 3 files under 4,500 words (about 6,000 tokens), because below that limit, Cowork reads everything at each session. Above it, it starts skimming and summarizing.
To save time, you can ask Claude to interview you to create these files. In a Cowork session, add the following prompt:
I'm setting up my space on Claude Cowork, build my a-propos.md file that gathers all the context to know about me and my activity. Your job is to interview me with 15-20 questions to understand who I am, how I work, and what I consider good work. Ask one question at a time. Then, compile the answers into a condensed markdown file of less than 2000 words.
Claude will ask you targeted questions, with clickable options. You answer in a few minutes, and you get a context file far richer than what you'd have written yourself. A good markdown file is worth more than 50 documents uploaded in bulk.
Adding, removing or switching the connected folder
You connect a folder from the desktop app, when starting a task or in a project's settings. Claude reads and writes there, and nowhere else. The Claude documentation is explicit about deletion, Claude asks for your permission before permanently removing a file. A prompt appears, and you click Allow.
To remove a folder, you disconnect it in the same place, in the session or project settings. The files stay on your disk. Archiving a project removes its metadata from the interface without touching the local folder. The right habit is to switch folders rather than add a second one. Every open folder widens what Claude can read.
Deliverables land where you asked for them. If your RESULTS folder is connected, Claude drops the finished .docx or .xlsx there. Otherwise the file stays in the session, where you preview and download it. Name the output folder in the prompt, it saves you the search afterwards.
The global instructions
You can add global instructions on Cowork that Claude will read before each task. Accessible in Settings
Here's an example of instructions to copy-paste:
GLOBAL INSTRUCTIONS
BEFORE EACH TASK
Read all the files in CONTEXTE/. No task starts without this reading.
Never read RESULTATS/ or PROJETS/ unless I point you to a specific file.
If the task involves a type of content that has a matching template in TEMPLATES/, first study the structure of that template. Use the structure. Do not copy the content.
FOLDER PROTOCOL
You have read folders and one write folder.
Read-only: never create, modify or delete anything here:
CONTEXTE/ → My identity, my writing rules, everything to know about me and my company.
PROJETS/ → My briefs, references and finalized work, organized by project.
Write folder: the only place where you deliver your work:
RESULTATS/ → Everything you create goes here. Save all deliverables under a subfolder named after the project. Create the subfolder if it doesn't exist yet.
NAMING CONVENTION
All created files must follow this format: project_content-type_v1.ext
Content types: Newsletter, LinkedIn Post, Brief, Deck, Report. Increment the version number if a file with the same name already exists.
OPERATING RULES
If the brief is vague or incomplete, ask me questions before acting. Don't fill the gaps with generic content.
Deliver the work. No superfluous commentary unless I ask for it.
Never delete files, anywhere.
The key point here: always prompt Claude to understand before executing. As you would with a new colleague, you want it to propose an action plan, identify what it's missing, and validate with you before rushing ahead. It hugely changes the quality of the result.
Concretely, in your prompts, you can add something like “Ask me questions to refine the approach”. This lets you have interactive questions with clickable options.
Global, project or folder instructions, which one applies?
The three levels coexist and do not replace each other. Global instructions live in Settings then General, under Instructions for Claude, and apply to every conversation, Cowork as well as chat. Project instructions only apply to that project's tasks. Folder instructions travel with a local folder, and Claude can update them itself during a session.
| Level | Where you set it | Scope |
|---|---|---|
| Global | Settings, then General | Every conversation, Cowork as well as chat |
| Project | Inside the project | That project's tasks, along with its memory |
| Folder | The local folder, in the desktop app | The selected folder |
Memory follows the same logic. It stays attached to the project and never crosses its boundary, so what Claude learns about one client does not resurface with the next. On the wider progression, from the chatbot that answers to the tool that executes, I laid out the steps in my guide to mastering Claude.
Skills = your reusable playbooks
It's an important concept, which I cover in my article on Claude Skills and Projects, so I'll go over it quickly.
A skill is an instruction file that tells Claude exactly how to handle a type of task. A kind of SOP for the AI.
Cowork has native skills for common files: Excel, PowerPoint, Word, PDF. When you ask for a spreadsheet, Claude automatically activates the Excel skill. Clean formulas, formatting, charts, not just a data dump.
On top of that, you need to create your own custom skills to guide Claude on the rules to follow depending on the tasks.
If it's the first time you run a specific task in Claude Cowork and it won't be the last, ask it to create a skill for that task.
Next time, the same type of task will be executed with the same quality, without re-explaining. The difference between a generic output and an output that sounds like your work is the skill.
Here's an example on generating a sales proposal (you can even see the assets used):
To manage your skills, go to Customize in the left sidebar, then Skills. You can enable, disable, import skills, or create new ones.
Plugins: ready-to-use business skills
Plugins are comparable to skills. Anthropic released 11 official ones at launch, in late January 2026, and its GitHub repository holds more than fifteen as of October 2026 (marketing, sales, legal, finance, data…). Each plugin gives Claude workflows and commands specific to a profession.
They're accessible in the same place as skills and once installed, you can trigger them by typing “/” in the chat to see the available commands. For example: /marketing:draft-content for a LinkedIn post or /data:explore-data to analyze a CSV.
Projects in Cowork
You've configured your space but the problem is that without a project, each session starts from scratch. There's no memory between sessions EXCEPT if you switch into a project.
Projects in Cowork change this. A project is a persistent workspace with:
- Its own folder (Claude reads and writes in it)
- Custom instructions (tone, rules, format, guardrails)
- Scoped memory (Claude remembers what it did in this project, from one session to the next)
- Scheduled tasks (recurring work, tied to this project)
- A history of all the tasks executed
Scoped memory is the feature that deserves your attention. You say “pick up last week's report” and Claude knows what you're talking about. But only within this project. Your “Newsletter” project knows nothing about your “Sales proposals” project. It's deliberate isolation, and it's a good thing.
How to create a project
In the Claude Desktop app, go to Projects in the left sidebar, then click +. Three options:
1. Start from scratch, create a new local folder and add your instructions and files
2. Import from a Claude Chat project if you already have a project in the browser, the files and instructions transfer over
3. Use an existing folder if you already have a well-structured COWORK folder, you wrap it in a project to add memory and instructions.
Once the project is created, you can launch a first task so Claude “learns” the context of the space:
I just created this project. Read all the files in the folder. Then summarize what you know about this space: what's there, what it's probably for, and the instructions you'll follow. If anything is unclear, ask me questions.
Claude will then have your project's context from day one and as long as you stay in the project. Here's an example on a “LinkedIn” project with the instructions and the context:
At Tandem, for example, we use a dedicated project for writing sales proposals. The folder contains the proposal template, 2-3 examples of proposals that worked well, the current client brief, and the call transcripts.
I drop a new client brief, I open the project, I type:
New client brief in the folder. Create a sales proposal in .docx. Ask me questions first.
Claude reads the brief, compares it to my examples, asks me 3-4 questions (”Budget? Decision timeline? Competitors in the running?”), and produces a .docx directly in my folder. Claude increasingly knows my formatting preferences and the level of detail I expect, even as my sales proposals evolve. It gets better with each session.
Scheduling, Dispatch and Team
Scheduled tasks
You've probably noticed, you can create recurring tasks in Cowork (including within projects).
Very relevant for monitoring, for example, or a morning routine that retrieves your unread emails (via connectors) and other notifications, your calendar of the day's events to generate a concise recap of urgent tasks, and saves it all into a morning-brief.md file.
That's at least 30 minutes of prep work done automatically before I even sit down at my desk.
3 points to keep in mind:
- A task that writes into a local folder needs the desktop app open. Since the move to the cloud, a task that only uses your connectors runs without your machine.
- Each scheduled task consumes your token capacity. Don't schedule useless tasks.
- Always test manually first. Run the prompt, check the result, adjust, and once satisfied, schedule it.
Dispatch, your phone as a remote control
Dispatch is a feature launched in March 2026 that lets you send a task to Claude from your phone. The workflow runs on your computer, and you receive a notification when it's done. Anthropic has since closed it to new users, and as of October 2026 only existing users can still use it.
The setup takes 2 minutes to sync Claude Cowork and your phone, but I'm not super convinced of the value for now. It can be worth it for long tasks (more than 15 or 30 minutes), but for runs of a few minutes, the value is more limited.
Cowork for teams (Enterprise)
Claude Cowork is available on all paid plans, with controls designed for organization-wide deployment. On Enterprise, an admin keeps control over whether it is switched on.
What's interesting in Anthropic's early feedback is that the majority of Cowork usage comes from outside the tech teams. Operations, marketing, finance and legal teams use Cowork not for their core work, but for everything around it: project updates, collaboration decks, research sprints.
An example of execution on the sales and finance side:
With results that can range from presentations (PPT), simulations (Excel), interfaces (artifacts), reports (docx) and more:
On the enterprise features side, we find:
- Role-based access controls where admins organize users into groups and define who can use which Claude features. You can enable Cowork for certain teams only.
- Per-group budgets. Each team has a spending cap, adjustable from the admin console.
- Usage analytics on Cowork sessions, active users, the skills and connectors used, everything is visible in the admin dashboard.
When skills become shareable across a company, Cowork goes from an individual productivity tool to a shared infrastructure. A skill created by one person can be used by the whole team. If you'd like us to support you on deploying Claude in your organization, we'd be glad to show you what we've done (and our approach).
What changes as Cowork moves to the cloud
Claude Cowork ran only on your own machine for a long time. Anthropic announced that in October 2026 new tasks run in the cloud on the Pro and Max plans, and that the “Only on your computer” option disappears from the settings. The most visible consequence hits scheduled tasks, which no longer require the computer to be awake.
The desktop app stays essential as soon as a task touches your machine. It is needed to read and write in a connected folder, to drive the browser, to let Claude click on screen and for local connectors. A monitoring routine that only queries your connectors therefore runs without you. A task that drops a file into your folder waits for the app to be open.
What to lock down before connecting a folder
A connected folder is read and write access handed to a program that decides its own actions. Anthropic says so in its own safety documentation and recommends a dedicated working folder rather than broad access, keeping financial documents, credentials and personal records out of reach. The named risk is prompt injection, when hostile instructions hide inside content Claude reads in good faith.
The French state has taken a position on these tools. In its CERTFR-2026-ACT-016 alert of April 2026, CERT-FR names Claude Cowork and OpenClaw among autonomous personal assistants, recalls that they run commands, drive the browser and send emails, and requires mandatory human approval whenever a system command is considered. Tandem detailed the steps to take in its article on AI agent security in the enterprise.
Yet the default setting points the other way. On Team and Enterprise, permission for the automatic approval mode is open out of the box, while “Always allow” on write-capable connectors stays closed. An admin closes the first one from the organization settings, and it is the highest-value move on the list.
The other admin levers exist and see little use. Cloud execution is on by default on Team and closed on Enterprise until a custom role opens it. Network egress permissions sit under Capabilities then Code execution, and apply only to new sessions. Cowork sessions surface in the Compliance API, and tool calls stream to a SIEM over OpenTelemetry. On splitting the roles between who decides and who builds, see our article on AI governance in the enterprise.
As soon as Cowork shipped, I noted in my edition on desktop assistants that Anthropic was already urging caution, because the tool can delete a file and the protections against prompt injection were still being built. The advice has not changed. In my one-pager on Claude inside Excel, the same rule sits in the footer, do not open files from unknown sources and always review deliverables before sending them.
The limits of Claude Cowork
Before wrapping up, there are still some sticking points:
- Project sharing is limited to Team and Enterprise plans. You grant view or edit access there, and you can revoke it later. On Pro and Max, a project stays personal.
- A project built from an existing folder on your computer stays there, and never reaches your Claude account. Only projects created from scratch or imported from Claude follow your devices.
- High token consumption. A Cowork session is worth about twenty classic chats. The persistent memory and context are worth it, but you will hit your limits faster on the Pro plan at $20 a month. Save the heaviest model for deep work, and leave simple tasks to a faster one. The full plan breakdown sits in our article on Claude subscription pricing.
- The browser extension is still unstable. Cowork can theoretically control your browser, but it's slow, it sometimes cuts out in the middle of a task, and it burns tokens. I advise against relying on it for now.
- The desktop app remains the gateway for local work. It has to be open to read and write in a folder, drive the browser or click on your behalf. If you want to roll out this kind of setup across a whole team, that is exactly what our Claude training covers.
Should you set up Claude Cowork in your company?
Yes, provided you treat the setup as a proper project. The gain is decided in the context folder. Twenty minutes spent on your three markdown files improve the deliverables more than any long prompt.
The order that works has three steps. A dedicated folder with CONTEXT, PROJECTS and RESULTS. Then one project per recurring topic, with its own instructions and memory. Then scheduled tasks, once the prompt has been proven by hand. Set the approval controls before the first sensitive task, not after. For heavy processing, on large files or dozens of exports to consolidate, Claude Code takes over.
Tandem has been rolling out this foundation with its clients since Cowork shipped, from an investment fund to a franchise network. If you want to set it up for a whole team, that is what our Claude training is for. And if you first need to work out which tasks are worth delegating, that is the job of an AI audit.
