Original title: I want to master Claude Cowork (full course). Original author: @hooeem Translated by: Peggy, BlockBeats
Original author: Rhythm BlockBeats
Original source: https://x.com/hooeem/status/2039723470691451072
Reprinted from: Mars Finance
Editor's note: This is a systematic practical tutorial focusing on Claude Cowork. The core is not about 'how to write better prompts', but about 'how to turn AI into an executable unit that can be managed and scheduled.'
The article starts from the basics and gradually unfolds a complete usage framework: first, establish a stable context through Project and .md files (i.e., the 'business brain'), and then achieve behavioral constraints through a three-layer instruction system (global, Cowork, project level); based on this, introduce connectors, browser, and desktop control capabilities to expand the execution boundaries of AI; ultimately, build a reusable and automatically running workflow system through Skills, Plugins, scheduled tasks, and Dispatch mode.
Unlike traditional 'conversational AI' usage, this method emphasizes structure over interaction, and systems over skills. Its goal is not to improve the efficiency of a single task but to crystallize repetitive work into long-term automatable processes.
If Chat is a tool, then Cowork is closer to a 'trainable digital employee'. This article provides a pathway to build from scratch and gradually amplify its capabilities.
The original text is as follows:
Claude Cowork launched three months ago, and since then, Claude has undergone over 50 updates, which has almost completely changed the way I use Cowork. Here’s how you can leverage these changes and truly master it.
Claude Cowork is changing the way we work. You just need to hand over tasks to it, and it will create plans, execute steps, and deliver completed results directly to your folder while you do other things. However, since Cowork was first released, many changes have occurred. This guide will clarify all the key content—so pay close attention.
If you want to hand over the complete workflow to a truly executable AI, then keep reading. This article is written for you.
Over the past few months, I have used it personally while studying how others use Cowork, and I have also consulted Claude to see if there are any interesting usage scenarios missed after updates.
In fact, I have polished this content into a complete course—whether you have never heard of Claude or use it every day, this article can help you truly master Claude Cowork.
What is Claude Cowork? (And what it is not.)
Claude Chat is an assistant: you input a prompt, and it provides an answer. You are always involved in the conversation. It is very suitable for brainstorming, drafting, and quick thinking—you are always in the 'loop'.
Claude Code is for developers: it runs in your computer's terminal, allowing you to write code, execute code, and manage GitHub repositories. It can also be wrapped in IDEs (like Google's Anti-Gravity) for use—it can be understood as a layer of Iron Man armor over 'raw intelligence.' But it is not designed for non-technical users.
Claude Cowork is an employee: it is an interface for non-technical users, but the underlying engine is the same automated execution engine used by Claude Code. You just need to assign tasks, and it will automatically break them down into subtasks, initiate a virtual machine environment locally, execute the entire process, and place the final results directly into your folder when you leave.
Chat is conversation. Cowork is task delegation.
The two are completely different relationships and will yield completely different results.
Once you truly understand this distinction, your perception of Cowork will change completely.
Setting method.
Cowork will read from and write to the files on your hard drive directly. Granting it access to the entire system often marks the beginning of a disaster.
So, first limit it to a controllable range. Do it like this:
Step One: Select a model.
On the right side of the Cowork interface, select the AI model to execute tasks.
· Sonnet 4.6: Main model. Costs less, is more efficient, and can cover 99% of your daily needs. Using it by default is sufficient.
· Opus 4.6: Heavy firepower. The consumption limit is quickly exhausted, reserved only for the most complex, critical high-risk tasks. To quote the 'Einstein principle'—don't let Albert Einstein do menial tasks in the kitchen.
· Haiku: suitable only for lightweight, speed-focused small tasks.
At the same time, remember to enable Extended Thinking at the bottom of the interface. It allows Claude to truly handle complex logic rather than just providing surface-level answers through pattern matching. The importance of this detail is greater than most people realize.
Step Two: Create a sandbox.
Right-click on the desktop, create a new folder named Claude Workspace or Sandbox. This is Cowork's workspace. The range of operations it can perform will only be limited to this specified folder. It cannot touch content outside of the folder.
Step three: Grant folder permissions.
After entering Cowork, click 'Work in a folder' and select the sandbox folder you just created. Claude will request permission to modify files within that location. Click 'Allow once' or 'Always allow' to proceed.
Now that the setup is complete, it's also more reassuring.
Step four: Execute your first task.
Drag dozens of different types of invoices into this sandbox folder, then enter the prompt: 'Please organize these invoices into different subfolders by category and generate a summary Excel sheet.'
Cowork will list the execution plan on the right side of the interface, then start processing independently. It may even schedule multiple parallel sub-agents to complete different parts of the task at the same time.
By the time you finish your coffee, the invoices may already be organized.
Build your AI workspace.
If you have been using Cowork in an 'unstructured' way, you may have encountered these problems: it does not remember anything between different sessions; every time you open a new window, you have to re-explain your business; different tasks interfere with each other and get mixed up.
There is only one solution—establish a project-based work system (project ecosystem).
What is a Project Ecosystem?
A project is not just a simple folder. It is a container that integrates the following: your files, custom instructions, skills, and context memory that can continuously accumulate... all concentrated in one place.
· What happens without a project? Claude does not know who you are, what business you are doing, or what tone you usually use. Thus, you will fall into a cycle: each conversation starts with explaining yourself from scratch.
· After having the project. Context begins to 'compound'. For example, in week 1, you need to write a whole paragraph to explain clearly; but by week 6, you only need to say 'do as usual', and Claude knows what you mean. This is the true power of Cowork.
A very important principle: projects must be isolated.
Different domains must be broken into different projects. For example: YouTube content project; finance project must not be mixed together.
What will happen otherwise? The tone and rules of YouTube will 'pollute' financial tasks; Claude will start to confuse styles, and the output will become unstable. This point is non-negotiable and must be strictly isolated.
Three ways to create a project.
You can create a project in three ways:
1. Start from scratch.
· Name the project.
· Write basic instructions.
· Continuously accumulate context during use.
Suitable for building a system from scratch.
2. Import from Claude Chat.
· Directly import your historical projects from the web version of Claude.
· All context and memory will be fully retained.
Suitable for users who already have accumulated experience.
3. Create based on existing folders.
· Choose a folder on your computer.
· Cowork will automatically build projects around these files.
Suitable for scenarios where materials are already available and quick initiation is required.
In summary: Cowork's capability lies not in 'single conversations' but in 'the long-term accumulation of project context.'
Your 'business brain'.
The project is the infrastructure, while the real 'personality' exists in .md files. These are plain text files placed in your context folder, which Claude reads before each task execution. It is through these files that it behaves as if it has been collaborating with you for two years, rather than being a general-purpose assistant that just started.
about_me.md
Clearly write who you are, what your business does, who your clients are, how you make money, and what your current priorities are.
Claude will read this file every time—every single time.
brand_voice.md
Define your expression style: the tone you like, expressions you dislike, common sentence structures. You can directly paste your past real writing samples.
The purpose of this file is to ensure that the content you write no longer resembles 'a generic Claude'.
working_preferences.md
Define your working method: how tasks are managed, where files are saved, and what format outputs should be.
Don't write from scratch by yourself.
The simplest way is to let Claude ask you. You can directly say: 'Please ask me step by step and generate a set of business brain files based on my answers.' It only takes 15 minutes. But it can save you hours each week, and this is ongoing.
Global instructions vs project instructions.
There are three layers in total, each layer will gradually tighten the scope. Only after configuring these three layers will Cowork work stably.
Layer 1: Claude Personalisation (Global).
Click on your name → enter settings. This layer affects all scenarios in Chat, Code, and Cowork.
This is used to set your 'universal rules', for example:
· Avoid excessive bold usage.
· Prioritize using primary sources when conducting research rather than secondary aggregated content.
· Avoid vague and cautious expressions (no hedging language).
This is your underlying behavioral norm.
Layer 2: Cowork Global Instructions.
Path: Settings → Cowork Settings → Global Instructions. This layer only takes effect within Cowork but applies to all tasks.
Suitable for writing rules that must be uniformly implemented, for example:
· Specify date formats.
· File naming conventions (e.g., underscore_descriptive_name).
· When handling business-related tasks, always check the sandbox folder first.
The purpose is: to avoid repeating the same context in each task.
Layer 3: Project-specific instructions.
Only take effect within a single project.
For example, in your YouTube project:
· Use specific slide tools to generate video openings.
· Record all data in a fixed spreadsheet.
These rules should not affect your financial projects.
Core principle: strict isolation.
Different project rules must be isolated. Otherwise, there will be 'pollution': YouTube's content rules affecting financial tasks; the output style becomes chaotic, and the logic is misaligned. Once mixed, Claude will begin to 'lose track of who it is.'
Establish global habits, Cowork executes, and projects become professional.
Equip your AI with capabilities.
From here on, Cowork is no longer just a useful tool but transforms into an operator capable of autonomously executing tasks. We need to empower it to step out of the sandbox and interact with applications, web pages, and your desktop environment.
The capabilities are divided into three layers, used in this order: first connectors, then browser capabilities, and finally computer operations as a fallback.
Connectors and MCP.
Connectors (also known as MCP, Model Context Protocols) allow Claude to operate directly in the applications you use daily.
Not through copy-pasting, not through screenshots, but by executing operations directly in the application interface through authorized access. The built-in connectors support Google Drive, Notion, Slack, Gmail, GitHub, Figma, etc.
How to connect.
Click the plus sign in the chat interface, go to the connectors tab, browse or search for the tools you need, and then complete the authorization process in the browser.
Once the connection is complete, you can perform fine-grained permission control. For each connected application, you can set it to 'always allow', 'needs confirmation', or 'forbid access'. You can always have complete control over what it can and cannot do.
Dedicated connectors
Taking Gamma as an example: If Claude generates presentations without dedicated connectors, the results are usually mediocre, with suboptimal formatting and structure.
But when you turn on the Gamma connector, Cowork will automatically hand the content over to Gamma, which will generate presentations with clear structure and excellent visual design.
The core principle is: use the right tool for the right task. Do not let a general-purpose tool do the work that should be completed by specialized tools.
Advanced techniques (Expert Hacks).
Hack 1: Apify MCP (painless data scraping).
Want to scrape data from YouTube, TikTok, or Instagram? You don't need n8n, nor do you need to set up complex automation processes. I have tried it before, and the difference is very noticeable in comparison.
Steps to operate:
· Create a free API token in your Apify account.
· Open Cowork's connectors, search for 'Apify'.
· Paste your API Key to complete the connection.
After completion, when you ask Claude to scrape data from a platform, it will automatically choose the most suitable one from over 1300 crawling tools from Apify (which they call 'actors') to perform the task.
You don't need to manually select or configure. It will automatically find the tools, run them, and return data.
Hack 2: Zapier MCP (connect 8000+ applications).
If the tool you are using is not natively supported by Claude, this is the solution. This is a feature that truly amazed me the first time I used it.
Zapier's MCP can connect over 8000 applications, such as HubSpot, Skool, Airtable, etc.
Steps to operate:
· Create an MCP server on the Zapier platform.
· Choose 'Claude Cowork'
· Configure the specific tools and operations you allow to be used.
· Copy the generated URL.
· Go back to Claude's connectors, search for 'Zapier', and paste the URL.
In less than 10 minutes, you can unlock 8000+ applications.
This entire integration directly eliminates a common problem: 'What if the tools I use are not supported?'
Now, this issue can basically disappear permanently.
Browser capabilities: autonomous web operations.
If no connectors are available, use a browser extension. This is the second layer of capability.
Installation method.
· Open Google Chrome.
· Go to the Chrome Web Store.
· Search for 'Claude'.
· Install and pin to the toolbar.
· Enable 'Claude in Chrome' in Cowork settings.
Post-activation capabilities.
Once enabled, Claude can:
· Automatically open browser tabs.
· Access the website.
· Read page content.
You can give it a URL to:
· Review the landing page.
· Analyze CTA (call-to-action button) design.
· Conduct competitive analysis.
It can not only read text but also understand page structure, such as:
· Italic emphasis.
· Title hierarchy.
· Page layout.
And it can perform multi-step operations.
For example, you can have it: open YouTube → browse recommended videos → return play count, like ratio, and comment sentiment.
It can execute completely.
A risk warning that must be taken seriously.
Claude uses your real browser. This means it is logged into your personal account. If you ask it to search for flight tickets, and you have already saved a payment method on the airline's website—technically, it has the capability to complete the payment directly.
So you need to:
· Monitor its operations.
· Set access restrictions (blocklists).
This is not a theoretical risk but a real capability. Treat it as a newly hired employee who already has access to your account.
Computer operation capabilities: full-screen and desktop control.
When connectors can't solve it and browser capabilities can't either, use Computer Use. This is the last resort.
It can allow Claude to:
· Directly 'see' your screen.
· Control your mouse.
· Input on your keyboard.
In other words, things you can do manually on your computer, it can now do autonomously.
How to get started.
· Click on your account name.
· Go to Settings → General.
· Open 'Computer Use'
Important: before performing any operations, add sensitive applications to the blocklist.
Actual usage scenarios.
For example, you can have Claude: find a video file on the desktop and drag it into the CapCut project.
The execution process is:
· Claude requests access to Finder and CapCut.
· You can operate completely without using the keyboard.
· It will visually recognize your files.
· Open the editing software.
· Find the corresponding file and complete the drag-in.
The entire process is completed automatically. This capability is very powerful. For this reason, you must prioritize setting up the blocklist.
Automation and task delegation.
There is a rule that will completely change the way you use this tool: any task that repeats more than once a week should be automated.
The automation system is divided into four layers: Skills, Plugins, Scheduled Tasks, Dispatch Mode. They are layered capabilities that do not need to be built all at once but rather constructed step by step in order.
Custom Skills: Compress workflows into a single command.
A skill is essentially a reusable AI workflow encapsulated into a single command.
You no longer need to write a long prompt every time to execute the same process.
Instead, you only need to call it with a slash command, for example: /summarise_invoices.
What is behind Skill?
Essentially, it is a .md file:
· Include instructions.
· Sometimes includes scripts.
· Claude dynamically loads while executing tasks.
You don’t need to write code. Just describe in natural language, and Claude will help you generate this file.
Skill creation process.
1. Definition (The Prompt)
Clearly describe what this Skill is supposed to do.
For example:
· Automatically apply your brand colors and fonts to all generated presentations.
· Organize a batch of invoices into a structured table.
2. Evaluation Loop.
Cowork has a built-in Skill testing mechanism:
· Execute a task using Skill.
· Do not execute once more using Skill (as a baseline).
· Compare two results side by side.
3. Iteration
Check results:
· Have any requirements been missed (e.g., forgetting to use a beige background)?
· Is the font correct?
· Does the format meet expectations?
Provide feedback to Claude, and it will automatically correct itself.
4. Save & Deploy.
When the output meets your standards: click 'copy to your skills'.
From now on, this entire process can be called with a single command.
The types of Skills you can build.
· Automatically parse a whole folder of invoices and generate a categorized Excel sheet.
· Capture YouTube videos and generate timestamped interactive HTML transcriptions.
· Call an external image model API to generate brand-style infographics with one click.
In summary: Skill = turning 'complex processes' into 'one action'.
Plugins: string multiple Skills together into a complete workflow.
A Skill can only automate one process, while the role of a Plugin is to string multiple Skills and connectors together to automate the work of a complete 'position.'
How to determine whether to use Skill or Plugin?
You can use a simple test:
· Is this task repeated every week? If not, just use a regular prompt.
· Does it involve more than 3 steps and cross more than 2 tools (like Slack + Gmail + Notion)? If not, use Skill. If so, then it should use Plugin.
How are Plugins built?
You only need to tell Cowork: which entire workflow you want to automate.
It will automatically combine the required MCP (connectors) and corresponding Skills into a complete 'main process package', allowing a single command to run the entire pipeline.
The most important value of Plugins: reusability and distributability.
Plugins can be shared. You can package your best workflow (SOP) into a Plugin and distribute it for your team or community to use. When they execute tasks, the effect will be almost identical to what you would do yourself.
In other words: your way of working has been productized.
Scheduled tasks: run automation on time.
Now, Skills and Plugins can automatically trigger execution at set times. Even if you are not in front of the computer, they will run.
How to set up scheduled tasks.
· Enter the 'Scheduled' tab in the left sidebar.
· Click 'New Task'.
· Fill in the task name and description.
· Reference the corresponding skill or workflow in the prompt. For example: 'Use my flight search skill to find round-trip tickets to Spain for under £2400.'
· Set execution frequency: every hour, every day, every week, or specify date + time.
· Choose the corresponding project folder.
· Save.
One very important rule.
Scheduled tasks will only execute when your computer is on and the Cowork application is running. If you set a task for 9 AM but your computer is off, then the task will wait until you turn on the computer to run.
A key detail that is easily overlooked.
It’s advisable to adjust power settings to keep the device awake. This is the most common source of problems—many automations 'fail' not because the configuration is wrong, but because the device is not running.
Dispatch mode: Remote control using a mobile phone.
When you are not in front of the computer, but tasks need to be executed on the local machine, use Dispatch mode. You can send messages to Claude via your phone, and it will execute tasks on your desktop. The phone and computer share the same conversation thread, and the context will not be interrupted.
How to set up.
In desktop Cowork:
· Open settings, enable Dispatch.
· Open the 'Keep Awake' option in the Dispatch menu (very critical). This can prevent the computer from sleeping, avoiding inaccessible files.
· Allow browser actions and computer use.
How to use.
Send commands from your phone, for example: 'Scan my receipts folder and generate a data dashboard.'
Cowork will execute the entire process on your computer. Once you enable push notifications, you will be alerted as soon as the task is completed.
Master-level workflows: prioritize building these three systems.
The following three sets are the most worthwhile to do first, in this order.
1. Daily Morning Brief.
Let Claude organize the day's information before you start working. It can be set as a scheduled task and connect to your calendar and email. It will:
· Summarize the day's agenda.
· List emails that need to be processed.
· Query local weather.
· Fetch the latest news related to your industry.
· Even draft email replies for you to review and send in advance.
As soon as you sit down at your computer, everything is already prepared. This alone makes it worth the subscription fee.
2. Content Repurposing System.
Give Claude a YouTube link, and it will automatically complete:
· Extract video transcriptions.
· Organize content into a new Notion page.
· Automatically generate platform-specific copy for LinkedIn and X.
One input, three outputs. No manual operation required throughout.
3. Financial Reporting System.
Set up a task to run automatically every month.
You can:
· Give Claude access to your transaction records (no need to provide complete account information).
· Or directly provide a receipts folder.
It will automatically:
· Categorize expenses.
· Check income and expenses.
· Generate an interactive HTML financial dashboard (showing profit and loss).
Your accountant can directly use this organized report, and you hardly need to spend any time on it.
Token Management: How to avoid quickly exhausting the quota.
Every word you send to Claude and every word it reads from files will consume tokens. If not managed properly, you could exhaust your quota within a week.
I have seen people exhaust their maximum plan quota in just three days due to three avoidable mistakes. They are as follows:
Error One: Ignoring basic load.
Before you input any content, a portion of your context window is already occupied:
· System instructions.
· The tools currently in use.
· Enabled MCP connectors.
These will all consume tokens. The more connectors you have open, the faster the consumption. Only open what is truly needed for the current task.
Error Two: Context Rot.
If you continuously use the same conversation window for a long time, Claude will reload the entire conversation history each time. For example, if you just finished writing an email and then ask it to plan a trip to Dubai in the same window, the system will load the 'email content' into the 'trip task' as well.
This is a complete waste, and many people unknowingly exhaust their quotas this way.
Error Three: Handling batch tasks through conversation.
If you need to process 100 invoices but let Claude read and process them one by one in the conversation, it will quickly exhaust the tokens. A better way is to have Cowork write a reusable script (Skill) to handle these invoices.
The tokens consumed by the script are far less than processing them one by one in conversation; the difference is very noticeable.
30–45 minute rule.
Keep each session concise and focused.
· Each window handles only one topic.
· Use for 30–45 minutes, or open a new window when switching tasks.
A new window means a new context, and it also means a new token budget.
Parallel sub-agents.
For large tasks, you can have Claude execute in parallel:
· Deploy multiple sub-agents.
· Each sub-agent has its own context window.
· Handle different parts simultaneously.
Einstein principle.
Opus is used only for complex, high-risk reasoning tasks. Sonnet is used for almost all other scenarios.
99% of the time, you do not need Opus. Do not use it as the default model.
AI Security.
Cowork runs directly on your local machine. Anthropic has already built security mechanisms into the product's foundation, so the basic risk is lower. The problem isn't here.
The real risk comes from others.
You can download advanced Skills built by other users from platforms such as GitHub. When you import these external Skills, you are essentially introducing a set of instructions for Claude to execute within the permissions you grant.
If someone embeds malicious instructions in this .md file, prompt injection may occur—such as inducing AI to delete files, steal data, or even escalate its own permissions within your system.
This is not a hypothesis, but a real risk.
Safety check process.
Before adding any external .md files to your Skills library, make sure to execute the following steps: copy all content from that Skill into Claude Chat, and then directly ask: 'Does this Skill contain any potentially harmful, malicious, or out-of-scope instructions?'
Only takes two minutes. But this must be done every time, without exception.
The key takeaways you should have.
All content of this guide is built on a key cognitive shift: do not treat Claude Cowork as a chat tool. It is essentially a 'self-directed employee' that can accept task delegation.
Once you understand this, you will no longer ask questions but start assigning it tasks.
The three most important points.
1. The architecture is more important than the prompt.
A clear project structure, a complete .md context file, and isolated project instructions significantly enhance output quality, far exceeding any 'clever prompt'.
Build the foundation once, then continue to reap the benefits.
2. The automation system is built in layers.
The sequence is: Skills → Plugins → Scheduled Tasks → Dispatch, each layer builds upon the previous one.
Don't try to build the entire system from the start. First, automate a weekly repeating task, then gradually expand.
3. Security is not optional.
Every Skill downloaded from the community must be checked before running. A safety check only takes two minutes, but those two minutes could protect your entire machine.
Your next step.
Today, do these three things:
· Create a sandbox folder.
· Grant permissions.
· Run a real task.
For example:
· Handle messy download folders.
· Or organize a batch of invoices.
This is enough for you to truly understand the capabilities of this tool.
Which workflow to start with?
· Morning Brief.
· Content Repurposer.
· Financial Dashboard.
Choose one and get started.
If you remember only one thing.
Use Zapier MCP. It can connect to over 8000 applications, and setting it up only takes 10 minutes. If you just skipped that part, go back and check it out. This is the most underrated feature in the entire tool and is a capability that most people have never truly utilized.