How to set up Claude Cowork.

Forget the other Claude. Focus on this one:

I was wrong.

I wrote “Cowork” and “Cowork 2.0.” My most-shared work ever, across Substack, LinkedIn, and X. A total of 20 million views.

Do not read them.

To be totally fair, I was right about one thing: Claude will become bigger than ChatGPT because of the launch of Cowork in January.

I’ve been begging you to switch to Claude since December 21st 2025.

Cowork is the best thing to happen to AI since ChatGPT.

If you don’t code, you must be using Claude Cowork.

Since then, I have written hundreds of guides and two major articles. I shared how to set up Cowork, and that was the latest (outdated) version of it:

Until today, that’s how I thought Claude Cowork must be set up.

The setup is good…

Until you go to enterprises, and they need to collaborate.

I started giving workshops to enterprises. Teams with hundreds of people using Claude Cowork together. And my files + folder system broke.

How Anthropic wants us to use their AI is different from how people actually use it.

And that’s OK.

Because I’m not paid by Anthropic.

So I can say whatever I want on how to best set up Cowork for you + your team.

I spent countless hours updating my setup. This guide shows you exactly what & how. By the end of it, you’ll know how to start - and what to absolutely avoid - with Claude Cowork.

Before starting, I want you to do two things:

1. Save this guide & block 20 minutes this week to set up Cowork.

2. Send it to anyone who still hasn’t tried Cowork (or Claude).

---

Skip this if you already use Cowork.

Quick reminder on how to access Claude Cowork:

1. Go to claude.com/download. Download the app on your computer.

2. You must have a Pro account ($20/month). I pay for the $100/month plan.

3. Open the app. Click on the Cowork tab at the top between Chat & Code.

4. You can add Skills and start Projects on the left menu. More about it later.

5. Make sure to select “Opus” for complex tasks. It’s the default model.

6. Once Fable is available, use it. It’s (somehow) even better.

---

I. Cowork is many Claude.

Cowork is much better than the normal Claude.

In simple terms, why is Cowork better?

Because it answers with many Claude at once.

Here, Cowork reads the Excel, and plans how to best solve my question.

The equivalent of 3x Claude already worked on my prompt before giving me an answer. This one is searching the web.

After 8 minutes and many, many Claude sessions, it gave me a great answer. I only asked something once, and Cowork planned everything else.

Claude Cowork brings many Claude to life to answer your question.

You have to imagine separate AIs collaborating to find the solution.

Cowork can perform a task for long minutes, sometimes 30+ min.

How to set it up is key, though. And I spent months running Claude Cowork wrong. If you used my old guides, so did you.

First, you must forget about this:

---

II. Forget about files and folders.

OK, hold up. I was the files and folders guy.

I think half of Linkedin is using files and folders on Cowork because of me.

It was a solid advice… but I was wrong.

But I was wrong.

Files and folders suck.

I told you to set up your Claude Cowork like this, with a folder on your computer and files (about-me files to know about you, and output files for Cowork’s work):

I would make you create a Cowork folder with files inside.

And I told you to select your folder (with files) when you start a Cowork session.

But after months of doing it, I saw two major issues:

1. Folders are leaking. Even if I told my Cowork to “never check my output folder,” Claude would sometimes do it. And it poisons my context with outdated outputs. Annoying.

2. Files are impossible to maintain. You have a full-time job. Me too. I can’t spend hours every week updating my “about-me” file. It was nice at first, and then a painful bottleneck.

Instead, you must focus on two things with Cowork:

---

III. Skills + Projects. Nothing else.

Skill = a capability you call. Travels into any chat. Fun fact, it also travels into any AI (like Gemini, ChatGPT…).

Project = a place you go to. Files + instructions stay loaded, it remembers, and you can share it as one place to collaborate with your team.

This is how I work with my team. This is how I work with enterprise teams at GPC. Skills and Projects.

Let’s go over them one by one and combine them afterward.

A. What you need to know about Skills.

Skills are defined by how you use them, right inside any chatbox, with the slash command (for eg, “/linkedin-post”).

Here’s a skill in action:

You invoke it with the /command. I’ll show you later how to build one and share it with your team.

This specific skill is built on how I make Linkedin hooks myself.

I have a library of my favourite Claude Skills. And I give for free to new subscribers in my welcome email. If you are already a subscriber, leave a comment & I’ll dm you:

You understood how to use a skill.

But how do you create a skill? Quite simple.

1. Either you use the Claude skills /skill-creator.

   1. You answer the question Claude will ask you.

   2. You test the skill. Save it. And invoke it with / command.

2. Or you use my free tool I built called “makemyskill.com”.

   1. I made it to suit how I like to build Claude skills.

   2. I give it for free because it costs almost nothing to run it.

OK now in practice it looks like this:

Create a skill inside Claude with the command /skill-creator.

You answer the questions.

Cowork goes into the multi-Claude sessions in one. Just wait.

You can save it here. And it has the name to / invoke it. Here it’s /contract-drafting.

That was the first way to create a skill.

But my favorite way is this one, if you start from scratch:

Go to www.makemyskill.com. Say what you want.

Create a free account (there is no paid version).

Now my little tool is creating your Claude skill by scrapping the web much deeper than what Claude usually do.

Click on “Download skill” the big fat orange button.

Go to Claude Cowork > Customize > Skills.

Skills > Add > Create skill > Upload a skill.

Upload the one we just made.

You can now use it by calling it with / command.

You can go very broad with skills, giving them tons of data that goes with the prompt. Some of my favorite skills have hundreds of pages of posts, hooks, SOPs, that Claude processes every time for this specific skill.

I forgot about another way to create a skill. If you’re having a chat with Claude, you can ask it to “make it a skill”.

Step 1: Go to the top of your chat, on the name of the chat.

Press enter. It will create a skill you can save.

And Claude will create the skill for you. Just wait.

OK, that was skills. Now projects.

B. What you need to know about Projects.

For Projects, it’s here:

Left menu > Projects > New project.

Start from scratch.

Give it a name + Add files. I prefer leaving the instructions blank.

You can see your Projects limits at the top right.

Prompt it. Your Project will have the context of your file inside the chat.

Every chat inside a Project has the entire context of this Project.

For example, you uploaded all of your clients’ information, the email you both shared, what you agreed on, their favorite past campaign…

Claude will remember it on every chat. And it does not poison the other projects (or chats) outside of your Projects.

C. A skill or a project?

I have a simple test:
It’s a skill you can teach someone → Skill.
It’s a specific context (like a client, a campaign) → Project.

You must teach your team the difference, otherwise you will end up with hundreds of skills that should be a couple of projects, or dozens of Projects that should be shared Skills.

I consult companies of 50-500 employees with my team in NYC. Message me if you want a discovery call - that’s how I make money and keep this newsletter free:

D. Combining the two.

Yes, you can use skills inside projects.

And that’s the best way to use Claude.

1. Use skills of your most recurring capabilities (eg. make a contract).

2. Use it inside a project of your client (eg. with his information uploaded).

3. Connect it before to your Gmail/Outlook and Slack/Teams (to get more context).

It looks like this:

Cowork knows everything, even how I like to negotiate. Guess what? My team also has access to my skills and project, in THEIR Claude.

Here’s how to share Projects with your team:

1. Create Projects not with Cowork, but Chat.

2. Go to the top right, and make sure it’s shared.

3. Once done, start any task with Cowork. It’s weird, I know.

This Project is inside Chat.

You can see your Projects limits at the top right.

Click Share and make it accessible to everyone.

Click on this little button.

Cowork will open a chat with the project’s context.

That was for Projects.

Now here’s how to share Skills with your team:

Left menu > Customize.

Skills > Click on one.

Click on “Share” and make sure your workspace is selected.

A reminder that skills are triggered with / commands.

E. Not everything is perfect.

Some problems I hate about this setup:

1. Skills can auto-fire from their description. I wish Claude could turn this off. I want to activate the skill only when I invoke it.

2. Projects have a limit on the number of files you can upload. I wish for more (obviously).

Speaking of which: Claude has a lot of settings to turn on or off.

And you should turn most of them OFF.

Here’s how I set up my Claude:

---

IV. No settings is the best setting.

I used to build everything around the folder, with files inside.

It was literally poison to my Cowork.

This was my old setup.

Two big problems:

• No one likes to create folders, and maintain them.

• No one likes to create files, and maintain it.

I also believe something, and everyone says the opposite: you shouldn’t give too many instructions to an AI.

Once you give too much context to Claude, it falls into the same answers every single time. No more creativity.

It’s OK when you need to draft the same contract.

It’s not OK to brainstorm and find creative solutions.

If you give too much context to Claude, it’s going into circles.

So here’s the very simple setup I have today:

#1. No more folders. And I use Opus 4.8 (High) most of the time. Max effort when in need, or Fable-5 once it’s available.

#2: In my settings, the minimum possible. Don’t give any info.

#3: Same here. Keep everything turned off but Connectors.

You can find Connectors in the + to connect apps like Gmail.

And now I can ask questions to my Gmail, inside Cowork.

#5: Turn on these.

Once you turn it on, Claude makes cool interactive charts like this.

#6: Keep your Global Instructions free of context.

You have now completed a setup of basically… nothing. Everything is mostly turned off. So what should you do instead?

1. Skills for recurring tasks.

2. Projects for recurring projects.

3. AskUserQuestion to prompt better.

You don’t know the third one.

But it’s my favorite Claude trick:

---

V. This is still my #1 trick.

You don’t know how to prompt Claude.

Or any LLMs to be honest.

It takes time, practice, experience, understanding of how the technology works. It’s not meant to say the truth, for example, but answering (which leads to hallucination, but it can be solved).

This simple trick will fix most of your prompt problems:

Claude creates a form you just have to answer.

Once you answer, Claude will think more about how to reply.

And now you have a much, much better file.

The AskUserQuestion trick is infinite.

It’s not just the first prompt. You can follow up with it forever.

And Claude will keep interviewing you until you’ve given so much context it’s impossible to fail.

---

VI. You really don’t need folders.

I know you are pissed at me because you’ve heard me saying files and folders are so cool.

And I really believed it.

But then I checked how I use Claude, and I just don’t open Claude’s files.

Let me give you an example.

When I create a spreadsheet, an Excel, using Cowork, I always export it with one button to my Google Drive (with Google Sheet).

Same goes to every document (text, PDF, slides…).

Here’s how it works:

Step 1 - I like using this template to make a clean Excel.

Create an Excel spreadsheet from:
[DATA: file path, folder, or pasted data, best is to upload data].

### Purpose:
[Who uses this and what decision/task it supports — 1 sentence.]

### Sheets needed:
- “[Sheet name]”: [columns, what each row represents, any calculations or formulas]
- “[Sheet name]”: [e.g., summary with totals, pivot, charts — or remove if only one sheet]

### Formatting:
[Currency/date formats, conditional highlighting, frozen header row, totals row — pick what applies.]

Before building, list your top 10 assumptions so I can sanity-check them, then execute.

Here’s an example:

Create an Excel spreadsheet: Customer support capacity and cost model for a fast-growing B2B SaaS company, covering Q3 2026 – Q2 2027.

Purpose:
Leadership-ready model to decide when to hire support agents, where automation helps most, and how support costs scale as customer count grows.
Business context:

* Company sells project-management software to SMB and mid-market customers.
* Current customer base: 4,800 accounts, growing 8% month-over-month.
* Support channels: email, live chat, phone, and self-serve help center.
* Ticket volume depends on customer count, onboarding activity, product complexity, and seasonality.
* Make your own assumptions for tickets per account, channel mix, first-response time targets, handle time, agent productivity, automation deflection, salary costs, software costs, escalation rates, and customer satisfaction impact.
* Every assumption lives as a labeled, editable input on the Assumptions sheet — never hardcoded inside a downstream formula — so the COO can pressure-test staffing and automation decisions live.

Sheets needed:
* “Assumptions”: all editable inputs grouped by Customer Growth, Ticket Demand, Channel Mix, Agent Productivity, Automation, Cost Structure, and SLA Targets
* “Ticket Forecast”: monthly ticket volume by channel, including new-customer onboarding tickets, recurring support tickets, escalations, and deflected tickets
* “Capacity Plan”: required agent hours, available agent capacity, hiring needs, utilization %, backlog risk, and SLA coverage by month
* “Cost Model”: support labor cost, software/tooling cost, outsourcing cost, automation investment, cost per ticket, and cost per customer
* “SLA & CX”: first-response time, resolution time, backlog, CSAT estimate, escalation rate, and churn-risk indicator
* “Dashboard”: single-screen executive summary with KPI tiles for total tickets, required headcount, support cost/customer, SLA attainment, automation savings, and CSAT trend; include charts for ticket volume by channel, capacity gap, and support cost over time
* “Scenarios”: Base / Growth Surge / Automation-Heavy toggle that flexes customer growth, ticket volume per account, automation deflection, and agent productivity

Formatting:
USD currency with no decimals, percentages where relevant, frozen header rows, clean executive-friendly palette, conditional formatting for SLA misses and utilization above 85%, monthly columns grouped by quarter, totals row on every model sheet, and named ranges for major assumptions so formulas are readable.
Before building, list your top 10 assumptions so I can sanity-check them, then execute.

I just received a multi-tab Excel.

And the magic button is the Google Drive integration:

Click here to open your file inside Google (docx, slides, sheets…). Much better than having a file inside your computer.

It automatically created this Google Sheet. Perfect.

And if you need to edit it, you have two choices from here:

1. Go back to Cowork and ask what you want. And if you don’t know what to ask, ask Claude to ask you. Using the AskUserQuestion tool I told you about.

2. Or use the (free) ChatGPT extension inside Google Sheets. It’s here:

Go to the top → Extensions → ChatGPT (you might have to add it first).

It will open a sidebar. I like using the Heavy mode.

I tested every single ways to create spreadsheets with AI. Cowork is the best.

And it’s not even the only cool part of Cowork.

You can onboard anyone of your company much more easily with Claude:

---

VII. Onboard with Claude.

Go to your Claude app. Create a new project.

1. Add all of the necessary resources, usually the company’s mission, deck, some interviews of the CEOs or C-levels…

2. If you can add best practices from support, they usually cover everything about the business a new employee should know.

3. Invite them to the new Claude Project.

4. Create a skill on top of it that is called /answer-onboarding with how to use the data inside the Claude Project to answer anyone doing an onboarding.

5. Now as soon as a new recrue needs an answer, they start by going to the Claude Project, and do /answer-onboarding + ask the question.

6. Claude will give a great answer.

I have a rule in my company:

• Please ask me as many questions as possible. Failing is winning.

• Never ask me a question before giving me a potential answer.

• Find the potential answer by doing what I just said: project + skills invoking how to (potentially) answer the new person.

It forces people to go from “I’m new, I can’t think for myself” to “I work actively to find the answer, but I still need help”.

It’s very different.

After this you’ll want to run everything through Cowork.

Don’t.

---

VIII. When Cowork is the wrong tool.

You should use…

• Chat: nothing to build, nowhere to keep it, a question, a quick rewrite.

• Claude Code: the deliverable is software, code, something to build.

• Claude Cowork: everything document-shaped (decks, analysis, client work).

If you want to use this visual, please reference https://how-to-ai.guide.

Claude is extremely powerful.

So powerful that its biggest blocker is most of the time… you.

---

IX. You are slowing us down.

This is a typical Cowork session:

1. You type a prompt. 30 seconds.

2. Cowork reads your skills. 50 seconds. It generates a plan. 50 seconds. It asks you clarifying questions using AskUserQuestion. 50 seconds.

3. Now you answer those questions. You click some options. Fine.

But sometimes you need to type a custom answer (and custom answers are where the best outputs come from).

So you stop. Think. Type.

60 seconds per answer. Maybe 2 minutes.

Across 8 questions, that’s 8-15 minutes of you being the slow part.

But Cowork can read 100,000 words in 15 seconds. It can build a spreadsheet in 90 seconds. And it has to wait for you to type at 60 words per minute.

But you could be faster. Because you speak at 150 words per minute.

Side note ⚠️

I know it sounds ridiculous to optimize everything for speed.

But talking instead of typing has another massive benefit: it sounds natural. It’s you, your voice.

And your brain thinks very differently when it has to talk.

Did you realize how good your ideas are, your flow is, when you’re talking to a colleague about solving a problem?

We want the same flow state here.

How to set up Wispr flow.

Quick reminder for those who missed it: Wispr Flow is a dictation tool.

You hold a key, talk, release. Your words appear wherever your cursor is.

Anywhere on your computer. Including inside Cowork’s chat box.

What makes Wispr Flow different: its accuracy.

Near perfection, every single time.

Wispr Flow + Claude Cowork is the perfect match to counter a common problem when using AI: you steer the conversation, and make sure you are in flow state.

Instead of typing “I need a Linkedin post,” you start talking “I recently found out about… and I want to share more about… but first I need to make sure that… so maybe we should start covering… and end up with… as a conclusion”.

1. The initial prompt: I speak it.

I say this out loud. Wispr types it.

The point isn’t to be faster, but we end up giving much more context when we talk (we are yappers by nature), rather than with a lazy typed prompt.

And the more context, the better.

2. AskUserQuestion answers: I speak those too.

Cowork generates a form. Most options I just click.

But when I need to add context (”make it more direct, she’s a CEO who hates fluff, and reference the ROI data from the last call”), I dictate that instead of typing it.

I don’t self-edit while speaking. I dump my thinking.

Cowork figures out what matters.

3. Feedback and pivots: spoken.

When Cowork produces something that’s off, I used to type feedback like: “Tone is wrong. Make it less formal.”

Now I say: “The tone is too stiff. I want it to sound like I’m texting a friend who happens to run a 200-person company. Keep the data but make it casual. Only redo section 2.”

Spoken feedback is richer because I am a yapper.

How to download Wispr Flow for free.

1. Go to wispr.ai. Download Wispr Flow. Install it.

2. The free tier is capped at 2,000 words/week - perfect to know if you like it.

3. Choose your favorite keystroke to activate Wispr Flow. I personally love “Shift”, the little arrow below “Enter”.

4. Go to any app, hold your keystroke (like shift) and… talk.

5. You now type 4x faster than before. Congratulations.

6. You get a free month when you pay for my newsletter (you don’t have to).

There’s no “integration” to configure. Wispr types wherever you type.

Here’s how it looks:

My Cowork literally spawned 5 running agents. Mini-Claudes.

Now there might be a problem.

Youll end up using Cowork so much you will spend a lot of money.

So here is how to save money in Cowork:

---

X - How to save money in Cowork.

The $20 paid plan of Claude will give you credits (called tokens).

But you will use them up very fast.

At the scale of a company, saving tokens = saving thousands of dollars!

This section is about saving as many credits (= tokens) as possible.

1 - Restart your conversation. Don’t send a follow-up.

First, Claude doesn’t count messages. It counts tokens. Every time you send a message, Claude re-reads the entire conversation history.

Message 30 costs 31x more tokens than message 1.

When Cowork gets something wrong, you want to type “No, I meant...” and send another message. Don’t.

Every follow-up stacks on top of the full conversation history. Claude re-reads all of it every turn. At ~500 tokens per exchange, 20 messages burns 105K tokens. 30 messages burns 232K.

Instead: click “Restart the conversation from here” on a previous message (much higher ideally so you do save tokens).

At the bottom right of everyone of your prompt, you can restart from here.

In Cowork, you can also edit a previous message.

Click on Edit on your own prompt. It will delete whatever comes next.

And you can edit your prompt to start over (from there).

So when something goes wrong, edit your prompt with a better prompt or restart the conversation higher.

It’s the biggest hack!

2 - Start a fresh session every 20 messages.

Long conversations are token furnaces.

One developer tracked his usage and found 98.5% of tokens were spent re-reading history. Only 1.5% went to the actual output.

When a Cowork session gets long: ask Claude to summarize everything, copy it, start a new session, paste the summary as your first message.

There is no specific prompt, but I like this one:

You keep the context. You lose the bloat.

3 - Batch your tasks into one message.

Three separate prompts = three full context reloads.

One prompt with three tasks = one reload.

Instead of: “Summarize this article” then “List the main points” then “Suggest a headline” write: “Summarize this article, list the main points, and suggest a headline.”

Using Wispr Flow helps.

4 - Use Sonnet (not Opus) for quick tasks.

Grammar checks, brainstorming, formatting, short answers. Sonet handles all of this at a fraction of the cost. And technically, Haiku is even cheaper.

Save Opus + Extended thinking for the work that actually needs it. Drafts and simple tasks on Sonnet-Haiku free up 30-70% of your budget for deep work.

Click on the model picker.

What about “Mythos” or “Fable-5” level of Claude?

Fable-5, called a “Mythos-level,” is even better than Opus.

It is obviously for your hardest tasks possible.

I’ll be honest, I was spamming Fable-5 requests whenever I could.

But you know it’s the most token-intensive model.

So there is no right or wrong answer. I am paying $100 a month for my Claude, and I have never run into any problems. We don’t use it for coding, so we are using fewer tokens than typical people.

5 - Delete your ABOUT ME file.

I wrote extensively on building your “about-me” file.

Keep it. But turn it into a skill:

It’s actually very simple. Just do this and upload it.

And I can now invoke it whenever I want.

6 - Spread your work across the day.

This one is impossible to actually do in practice, but it works.

Claude uses a rolling 5-hour window. If you burn your entire limit in one morning session, most of your daily capacity goes unused.

Split into 2-3 sessions: morning, afternoon, evening.

By the time you come back, your previous usage has rolled off.

And avoid peak hours (5-11 AM Pacific on weekdays) when the same query costs more against your limit.

Cool in theory, but I never do it ahah.

If I want to work, I work.

That’s why I pay for the $100/month paid plan.

So far, no problem (but I had a lot on the $20 plan).

I think a couple of dollars per day is worth it to have a godlike intelligence by my side. Your call.

---

XI. Your first 20 minutes.

• 0–3: empty your settings, turn everything off.

• 3–8: build your first Skill (something you keep doing).

• 8–13: create your first Project (upload PDFs, excels… from a client).

• 13–17: use your skill inside your project to see how it feels.

• 17–20: download my Claude Skills library by subscribing for free.

I give the library for free in my welcome email.

It looks like this:

It’s all for free for new subscribers.

If you are already a subscriber, leave a comment & I’ll dm you:

Creating skills should be your priority once you set up Cowork.

If you want (yet) another way to create skills, here’s a cool last example:

I saw this tweet praising a lot of experts.

Go to Chat (not Cowork) and turn on “Research”. I then copied and pasted the tweet.

It generated a long research on the topic after a deep web search.

Claude decided to create 7 skills out of it.

Sky is the limit. Create many skills, test them, improve them.

I need to clear things out:

---

I’m not paid by Anthropic.

I don’t pick sides.

Not Claude, not anyone.

Nobody pays me to write this.

Twice a week I show you how my own work is changing with AI.

As fast as it’s actually changing.

I want to be your filter to the noise — the person who tells you what to ignore and what to set up this weekend.

750,000 of you trust me to be that filter. You probably know me from my Linkedin. Or someone you trust forwarded you one of my guides.

If this one helped, be that person for someone on your team.

Share it with them.

It costs you nothing, and it makes your close network better at it.

And if someone sent you this, thank them, then subscribe so you get the next one free.

---

PS: You don’t want your company to be left behind.

Reading about AI is useful.

But it won’t change how your company works.

If you have 100+ people on your team, adopting AI is not a “send them a newsletter” problem. It is a systems problem.

Your people have different roles, clients, tools, workflows, permissions, data, risks, and habits. My newsletter, sent to 750,000 people, can’t do that.

That is how I help companies.

I run tailored AI workshops for enterprise teams. I am not alone, we are 20 people in New York City to support our clients.

We map where AI can actually save time, where it should not be used, and how your people can adopt it without creating chaos.

If your company has 100+ people, message me.

My team & I will take care of you in priority.