> 📥 The full guide is available as a PDF (Danish, 9 pages). Download it at the top of the page — set aside an afternoon and you're up and running with Cowork as a real AI colleague.
1. What Cowork is, and why now
Claude Cowork is Claude Code's agent architecture wrapped into the desktop app so you skip the terminal. You point it at a folder and give it a task, and it plans, works in parallel tracks and delivers finished files: documents, spreadsheets with working formulas, presentations.
The mental shift is bigger than the technical one. With a chat you write one draft at a time. With an agent you describe a result and come back to finished work. Many people still sit and write emails with their AI license. Meanwhile the tools have moved up a level, to doing the work itself.
Setup is what decides whether Cowork becomes an AI colleague or just another chatbot. Cowork has no memory between ordinary sessions, so the context you build up in your folder IS the tool's memory.
> The most important principle: With a chat it's about what you type in the field. With an agent it's about what you give it access to. Decide access before you open the app.
2. Before you start
Only two things are required: the desktop app and a paid plan. The rest is good housekeeping.
- Claude Desktop app — macOS, Windows and Linux (beta), claude.com/download.
- Paid plan — Max for daily use. Team/Enterprise at work.
- Internet — stable connection throughout the session.
- Your computer — the app must stay open while Claude works.
All paid plans run full Cowork, but usage is the real difference. Cowork burns markedly more tokens than chat, so the Pro cap is hit quickly. Plan on a Max plan if Cowork is going to be part of your workday. And if you use it seriously at work, with company or client data, you belong on Team or Enterprise: the DPA and the no-model-training default come with those.
3. The 10 steps
1. Download Claude Desktop. Cowork lives in the app, not the browser. Get it for macOS, Windows or Linux (beta) at claude.com/download, and find the Cowork tab next to Chat.
2. Pick a paid plan. Cowork eats tokens. Pro opens the door, but the cap hits fast on daily use, so plan on Max. Serious work use = Team or Enterprise.
3. Create ONE dedicated folder. Make a Claude-Work folder and point Cowork at it. Never Desktop or Documents from day one. The folder is your safety boundary.
4. Fill the hub. AboutMe (who you are, how you write). Templates (your preferred structures). Project (pointers to the real work). Outputs (5–10 of your best deliverables — they set the quality bar).
5. Write CLAUDE.md. A short text file at the top of the folder: what the folder is, who you are, your voice, how folders are organized. 200 words is enough. Have Claude interview you and write it.
6. Set Global instructions. Settings > Cowork > Global instructions. The evergreen rules that apply every session.
7. Make an anti-AI-writing file. A list of words, phrases and constructions Claude may never use when writing in your name. Without it, everything sounds like a chatbot.
8. Start low-risk. First task: clean up Downloads, or build a deck from raw notes. Low risk, fast result. Keep client data out until the routine and subscription are in place.
9. Schedule recurring tasks. Write `/schedule` in a task — e.g. a news briefing delivered as a Gmail draft every morning. Only requires the computer to be awake and the app open.
10. Check everything. Always. Output is 95 percent perfect and 5 percent embarrassing. Cowork is strong when it moves and structures what exists, weak when it generates something new: names, headlines, details. Put your control where it generates.
4. What works and what doesn't
Cowork is reliable when it moves and structures existing material, unreliable when it generates new content. That distinction explains almost every success and every embarrassment I've had with it.
Works: folder cleanup and structure, decks from raw notes (80–90 % done), Excel analyses (talk to your data), scheduler for recurring work.
Disappoints: browser control, anything with 2FA and payments, self-invented names and headlines, memory outside projects.
> Check everything. Always. The 5 percent embarrassing hides in what you don't read carefully: an invented headline, a wrong name, a number that almost fits.
5. Delegate on the result
Vague instructions give vague results. Cowork works asynchronously: you hand off the task, it works, you come back. So your message must describe the end result and everything Claude can't guess. Answer three questions before you hit send: What does "done" look like? What context is required? What constraints apply?
Always ask for a change-log so you can audit each choice afterwards. And test on throwaway files first.
6. Data discipline, before you open the app
Set the frame once, before the app opens: Cowork gets its own working folder, and everything else stays out. Where the line goes depends on your subscription and your data.
7. Data residency, PII and GDPR
Cowork doesn't change the law, but it moves the risk. A chat only sees what you type. An agent reads whole folders. So three things must be in place before personal data goes near it: where the data is processed, whether it's used for training, and who's responsible for what.
Status as of July 2026:
- Data residency: No EU-only guarantee — data is processed in the US.
- Model training, Free/Pro/Max: Can occur, controlled by 'Help improve Claude'.
- Model training, Team/Enterprise: No training by default, DPA available.
- Sensitive data (national IDs, health): Keep out of Cowork entirely.
> Scrub before, not after. An anonymization tool can't clean a file after Claude has read it. Order: scrub, place in folder, grant access.
8. New in Cowork right now
- Projects: dedicated workspaces with their own files, instructions and memory.
- Mobile: give Claude tasks from your phone while the computer works at home (Pro and Max).
- Plugins: skills, connectors and sub-agents bundled to your role.
- Skills: build small, split skills rather than one big one.
> ⚠️ Disclaimer. General practical guidance — not legal or compliance advice. Always consult your own compliance or legal function and your DPO, and verify dates and status against primary sources (EUR-Lex, European Commission, Datatilsynet). Status as of July 2026.
📬 Subscribe to "AI, Built Human" on Substack — weekly insights on AI in practice.
Stefano Vincenti · GenAI strategist and architect · External lecturer, IT University of Copenhagen & DIS Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone · aitrainer.dk