Claude Cowork, set up right.

What you'll have by the end

• Claude Cowork installed and configured on your Mac or Windows PC
• Global instructions that make Claude work the way you work
• An about-me memory system Claude reads at the start of every session
• Connected tools: Gmail, Calendar, Drive and more
• The essential skills switched on, and a folder structure that scales
• One official plugin installed and briefed on your business
• Guardrails set, so Claude proposes and you approve

One mindset shift before you start. Treat Claude less like a tool and more like a capable colleague who uses your tools: files, inbox, calendar. The setup below is what you'd give any new team member: context on who you are, rules for how you work, and access to what they need. Nothing more.

This is the setup hour. Putting the setup to work (your first agent, scheduled tasks, real spreadsheets and decks) is what our hands-on course, Claude Cowork Essentials, covers; the last tab tells you how to go further. And if you'd like the broader orientation first, our free Getting Started with Claude guide covers the whole landscape: the web app, prompting, and where Cowork fits.

A note on accuracy: menus, models and prices move fast. Everything here was verified against Anthropic's official documentation and a live account on the date in the footer. Where the app in front of you disagrees with this page, trust the app.

1.1 · Get the app

No Claude account yet? Create one free at claude.ai first; the plan comes in 1.2. Cowork also runs on the web there, in beta, with nothing to install; it's a fine way to try it. Install the desktop app for the full setup in this guide: it's what lets Claude work in folders on your computer.

1.2 · Choose your plan

1.3 · Open Cowork

1. Stay on Chat Top-left, above the sidebar, is a Chat / Code toggle. Leave it on Chat. Code is for developers.
2. Flip the switch In the message box (“How can I help you today?”), flip the Chat | Cowork switch to Cowork. Hover over it and it reads “Give Claude a task to run autonomously.” (On iPhone, Cowork has its own sidebar item.) That's home for the rest of this guide.
3. Leave it on Manual Under the box you'll see a Project or folder picker and a permission setting. Ours started on Manual, which keeps you approving Claude's actions; leave it there while you learn.

1.4 · Create your workspace folder

1. Make the folder In your Documents folder, create a new folder called Claude Cowork. The same path works on Mac and Windows.
2. Connect it Click Project or folder under the message box, choose the folder, and allow access. On the merged version, granted folders are listed in Settings under Trusted folders.

Treat this folder as the desk you share with Claude. Everything you want Claude to reach goes inside; everything you don't, stays outside. That one habit does more for safety than any setting.

1.5 · Pick your model

The chip in the message box shows the model and the effort level (ours read Opus 5.5 · Medium); click it to change either. Start where it opens. Drop to Sonnet 5 when speed matters more than depth, and raise Effort, how long Claude thinks before acting, for a complex, multi-step job.

1. Copy the brief into a chat Copy the Cortexa starting brief below and paste it into a new chat. Claude asks for each {PLACEHOLDER} you haven't filled, then hands back the finished brief.
2. Answer in plain language Skip anything that doesn't fit how you work and Claude drops the line. Copy the completed brief it returns.
3. Open global instructions Click your name or initials (bottom-left), choose Settings, open the Cowork section and find Global instructions. Click Edit. On the merged version, this box has moved to General → Instructions for Claude.
4. Save Paste the finished brief into the window and click Save.

You're helping me finish my Claude Cowork global instructions.
Below is my starting brief. Ask me for each {PLACEHOLDER} that is
still unfilled, a couple at a time, in plain language. If I say a
line doesn't fit, drop it. When everything is answered, give me the
finished brief as one copyable block, no placeholders and no
commentary, ready to paste into my global instructions in
Settings.

---

# Who you're working with

My name is {FULL NAME}; call me {PREFERRED NAME}.
Role: {JOB TITLE} at {COMPANY OR TEAM}.
In one sentence, my work: {WHAT YOU DO AND FOR WHOM}.

The people my work serves: {CLIENTS OR AUDIENCE — who they are and
what they count on you for}.

My priorities right now:
- {PRIORITY 1 — name the outcome, not the task}
- {PRIORITY 2}
- {PRIORITY 3}

# How to work with me

- Answer first, context after. Don't repeat my question back to me.
- Plain language. If a term needs explaining, choose a plainer term.
- Give me the one thing I asked for. Offer extras in a single line,
  not as a second deliverable.
- If my request is vague, ask one clarifying question, then commit
  to your best reading of it.
- Default to prose in conversation. Save headers, bullets and tables
  for when the content genuinely needs them.
- Preferred format for written work: {MARKDOWN / PLAIN TEXT / DOC}.
- Unless I say otherwise, write for {WHO USUALLY READS YOUR WORK}.

# Candor

- Tell me what's real. If a plan has a flaw, say so before I act on
  it, not after.
- Disagree when you have a reason. I'd rather be corrected than
  flattered.
- Skip compliments about my questions or ideas; grade the ideas
  instead. Strong, workable or weak, and why.
- When you present options, rank them and name your pick.
- Separate fact from estimate. Label your guesses and say what would
  confirm them.
- If something won't work, lead with that, then offer the closest
  thing that would.

# Guardrails

- Confirm with me before anything hard to undo: sending, publishing,
  deleting, sharing outside this workspace, or spending money.
- Anything outbound (email, messages, posts) stays a draft until I
  send it myself.
- Client material is confidential. Never carry one client's files,
  data or voice into another client's work.
- Client-facing work gets my review before it ships. No exceptions.
- Cite real sources or say you don't have one. Never invent
  references, numbers or quotes.
- On legal, medical or financial questions, give me your read, then
  flag what a professional should confirm.

# When in doubt

1. Candid beats comfortable.
2. Plain beats polished.
3. Short beats long, unless I asked for depth.
4. Ask once, then act.
5. Finish the work before describing the work.

You don't write these by hand. Paste the prompt below into a new Cowork session and let Claude interview you and build all three.

Create an "about-me" folder in this workspace with three files
you'll read at the start of every session:

1. about-me.md — my name, role, business, the tools I use, the
   people my work serves, and my current projects. Write it as the
   briefing you'd want on day one of a new job.

2. writing-rules.md — rules for writing as me. Start from
   Wikipedia's "Signs of AI writing" guide and distill it into a
   short list of what to avoid, plus anything you notice about how
   I write.

3. memory.md — a running project log. Append new entries at the
   bottom, update old ones in place, and keep a short list of open
   loops so nothing drops between sessions.

Interview me for whatever you're missing, then build all three.
When you're done, draft an updated global-instructions block that
has you read these files at the start of every session, so I can
paste it into my Settings.

1. Answer its questions Claude will ask about your work before writing. The more specific your answers, the better every future session gets.
2. Update your settings Paste the instructions block Claude drafts into your global instructions, in the same place you used in Step 2.

Before we start anything: how are you set up to work with me?

If the answer mentions your three files, the system is live. Every session from now on starts already knowing you.

This is the part of the setup that compounds. Ten minutes of instructions improve every session equally; the memory files get better every session, because Claude keeps adding to them. A month in, memory.md knows which arguments landed with which client and which drafts died on the vine, and you stop paying the re-explaining tax at the start of every conversation.

4.1 · Add your connectors

Open Customize in the left sidebar (it gathers connectors, skills and plugins in one place) and add connectors in this order. The + in the message box has a Connectors menu too, for switching them on and off as you work.

4.2 · Decide on the training setting

In Settings → Privacy, review the Help improve Claude option. Left on, your sessions can be used to improve Anthropic's models and are kept, de-identified, for up to five years; switched off, they aren't used, and the standard 30-day retention applies. The setting applies to Free, Pro and Max accounts, so set it deliberately, the way you and, if relevant, your clients would expect. Team and Enterprise accounts fall under Anthropic's commercial terms, and their content isn't used for training.

4.3 · Test the connection

Quick connector check: look at tomorrow on my calendar and at
today's unread email, and tell me the three things that matter
most. Flag anything you couldn't reach.

4.4 · Set the draft-only email rule

The Gmail connector can now send, reply and forward, and by default it asks for your approval before each one. Before Claude touches your inbox in earnest, add one standing rule on top of that: it drafts, you send.

Standing rule for email: anything you write goes to my Drafts
folder for review. You never send. Record this in memory.md so it
holds in every future session.

4.5 · Map your workspace (if your notes live elsewhere)

If your working knowledge lives in a big space (a Notion workspace, an Obsidian vault, a deep Google Drive or SharePoint), a one-off map saves Claude from searching the whole thing every time it needs one folder. It's ten minutes that pays back on every lookup after it.

A note for Obsidian, Logseq and similar local-file note apps: they need no connector at all. Your vault is an ordinary folder of Markdown files, so connect it the same way you connected the workspace folder in Step 1, and Claude can read it directly.

Go through my {NOTION WORKSPACE / OBSIDIAN VAULT / GOOGLE DRIVE /
OTHER} and build a map of it: the key databases or folders, the
active projects, and where reference material lives. If I left that
placeholder unfilled, ask me where my notes live before you start.
Save the map as my-workspace-map.md in the about-me folder, and
note in memory.md to check the map before searching from scratch.

4.6 · Optional: Claude in your browser

The Claude in Chrome extension lets Claude browse sites, fill forms and click through workflows in your own browser (on Max and Team, and rolling out to Pro). It's genuinely useful and genuinely powerful, which is exactly why we suggest adding it after the basics feel comfortable, not on day one.

5.1 · Enable the essentials

1. Open the directory Go to Customize in the left sidebar, open Skills, click + and choose Browse skills. The + in the message box has the same Skills menu, with Manage skills and Browse skills.
2. Check Skill Creator Find Skill Creator in your skills or in the directory, and make sure it's on. It's the skill that helps Claude build other skills, and the course leans on it heavily later.
3. Add what you'll use Click Add on the document, spreadsheet and presentation skills that match your actual work; added skills switch on right away. Skip the rest, since adding one later takes seconds.

5.2 · Build the folder structure

Set up the rest of this workspace:

1. A "projects" folder. Each active project gets its own subfolder
   containing a brief.md (what the project is and what done looks
   like) and its own memory.md.

2. An "outputs" folder. Everything you produce lands here, in a
   subfolder per project, so I always know where finished work is.

The about-me folder stays as it is. If I bring you something new
and you're not sure whether it's a project, ask. Then record this
structure in memory.md so it holds.

Claude Cowork/
├── about-me/
│   ├── about-me.md
│   ├── writing-rules.md
│   ├── memory.md
│   └── my-workspace-map.md
├── projects/
│   └── [project-name]/
│       ├── brief.md
│       └── memory.md
└── outputs/
    └── [project-name]/
        └── [finished work]

The shape matters more than the names: context in one place, live projects in another, finished work in a third. Claude keeps it tidy because the memory file now tells it to.

It earns its keep three ways. A brief and a memory file per project mean work resumes mid-thought instead of being re-briefed: pick a project back up after two weeks and Claude already knows where it stood and what's still open. Keeping each client in their own folder is the guardrail from Step 2 made physical, so one client's material can't drift into another client's work. And an outputs folder means finished work is never “somewhere in the chat”: it's a file, in a place you chose, ready to send.

6.1 · Install one official plugin

If a skill is a tool in Claude's toolbox, a plugin is a specialist who arrives with their own toolbox: methods, checklists and knowledge for a whole domain. One is plenty for now, and Anthropic's own are the safe pick.

1. Browse the directory Go to Customize, open the Plugins tab and look at what's available.
2. Pick the official one that fits Claude for Small Business covers finance, sales and marketing workflows; the Claude for Legal practice-area plugins cover contract triage and meeting prep. Choose whichever maps to your actual role.
3. Brief it A specialist who knows your business beats a generic one. Paste the prompt below into a new session after installing.

You now have the {PLUGIN NAME} plugin. (If that placeholder is
still unfilled, ask me which plugin I installed first.) Read my
about-me files, then tell me: which three of its workflows would
help my work most, and what would you need from me to run the
first one? Record the answer in memory.md.

6.2 · Scheduled tasks, at a glance

Scheduled tasks are the same idea as everything above, on a cadence: one standing brief that runs daily or weekly without being asked twice. They run in Anthropic's cloud, so a closed laptop doesn't stop them; only tasks that need your local folders wait for the desktop app to be open. When you're ready, typing /schedule in Cowork starts one, on any paid plan, and Scheduled in the left sidebar lists them all.

Your safety net, layer by layer

1. Folder scope Claude reaches only the folders you've connected. The shared-desk habit from Step 1 is the foundation everything else stands on.
2. Manual permissions With the setting under the message box on Manual, Claude asks before it acts. Loosen it only once you know how it works.
3. Standing rules Your starting brief requires confirmation before anything hard to reverse: sending, publishing, deleting, spending.
4. Draft-only email The Gmail connector can send, and asks your approval first by default. Your standing rule adds a second lock: messages stay in your drafts folder until you send them yourself.
5. The training setting On a personal plan, you chose in Step 4 whether your sessions may help improve Anthropic's models. Team and Enterprise content isn't used for training.
6. Connector permissions Each connector has its own controls, and you connected the minimum. What isn't connected can't be read.
7. Computer use If you later enable computer use, review which apps Claude may operate before you rely on it. Keep it away from anything you'd hesitate to hand a contractor.

Working with regulated or genuinely sensitive data? That's a conversation with your compliance team before it's a Cowork task. On Enterprise, cloud Cowork sessions are captured by the Compliance API, while sessions run on your own computer keep their history on that machine, where admins can't manage it centrally. Ask your admin what's recorded before assuming the trail exists.

# Claude Cowork setup checklist

## Step 1 — Workspace
- [ ] Claude Desktop installed (Mac: drag to Applications · Windows: run installer as admin)
- [ ] Windows only: Virtual Machine Platform enabled, then restarted
- [ ] Paid plan active (Pro or higher; Cowork isn't on Free)
- [ ] Cowork switched on in the message box, permissions left on Manual
- [ ] Workspace folder created and connected (Documents/Claude Cowork)
- [ ] Model and effort checked (start where it opens; more effort for hard jobs)

## Step 2 — Global instructions
- [ ] Starting brief filled in and saved in Settings

## Step 3 — About-me memory system
- [ ] about-me.md created
- [ ] writing-rules.md created
- [ ] memory.md created
- [ ] Global instructions reference all three files

## Step 4 — Connectors
- [ ] Gmail connected, draft-only rule recorded
- [ ] Calendar connected
- [ ] Drive connected
- [ ] Workspace map built (if your notes live in Notion, Obsidian, Drive or similar)
- [ ] Training setting reviewed (Settings → Privacy; personal plans)
- [ ] Claude in Chrome installed (optional, later)

## Step 5 — Skills & folders
- [ ] Skill Creator on, plus the document/spreadsheet/deck skills you'll use
- [ ] Folder structure built (about-me / projects / outputs)

## Step 6 — Extras
- [ ] One official plugin installed and briefed
- [ ] Read the scheduled-tasks overview (the build itself comes with the course)

Six habits that compound

• Dictate long prompts instead of typing them: built-in dictation on the Mac, Win+H on Windows, or a tool like Wispr Flow.
• Once your starting brief is solid, leave it alone. Day-to-day changes belong in the about-me files.
• A workspace map saves time and tokens: Claude finds things instead of searching for them.
• When you discover a better workflow, tell Claude to record it in memory.md. That's how the setup keeps getting better.
• Watch the progress panel on early runs. You'll learn how Claude thinks, and catch wrong turns while they're cheap.
• Match the model to the job: Sonnet or Haiku for quick work, Opus or more effort when the stakes are high. It stretches a Pro plan a long way.

Specific in, specific out. The quality of everything above tracks the quality of your about-me files: vague context produces vague work, from AI and from colleagues alike. If writing the first ask is the hard part, our free Prompt Generator matches you three ready-to-run Claude workflows in about two minutes.

Go further: Claude Cowork Essentials

This guide ends where the interesting part begins. Your account now knows you, reaches your tools and keeps its memory; what it hasn't done yet is real work. That's the ground Claude Cowork Essentials covers: Cortexa Academy's hands-on course where you build your first agent, put jobs on a schedule that runs while you sleep, and turn real files into spreadsheets, decks and dashboards, with the patterns worth stealing along the way. Intelligence, applied.