I about Claude Code and how you can use it to automate processes and be more efficient as an engineer. However, Anthropic also has a tool called , which you can access either through their web interface or their application.
Cowork typically targets less technical people with a simpler UI while still keeping a lot of the capabilities that you have in Claude code. In this article, I’ll discuss how you can get the most out of Claude Cowork. Of course, this is very useful to non-technical people who don’t want to interact with a coding agent through a terminal, but I also believe Claude Cowork can be useful for technical users with features such as
- A cleaner interface
- Visualizing figures directly in the app
- More focus on interactive UI elements
In this article, I’ll cover on a high level how to get the most out of Claude Cowork with some specific techniques that I utilize.
I’m not sponsored by Anthropic in writing this article; I’m simply a user of their product.
Why use Claude Cowork
I think there are two main sides to why you should use Claude Cowork. One side is if you’re a non-technical user. If you don’t work much with coding agents, Claude Cowork is a great introduction. It simplifies the interaction a lot, where you simply interact with Claude through a web interface or Mac/Windows application.
Cowork is, in many ways, just a simpler way to access the powerful features of Claude Code. You get all the same capabilities and options as you get when you interact with Claude Code through a CLI, but it’s presented to you in a nicer manner, and it makes it easy to get an overview of all the tasks you’re working on.
A lot of people might think Claude Cowork is completely useless if you’re a technical user and are already using Claude Code through your terminal. However, I would argue the opposite, and I do find myself using Claude Cowork at times where I need more focus on the visual experience, or I have separate, less technical tasks that I don’t want to perform via the terminal.
How to effectively use Claude Cowork
Now, let’s move on to how you should be using Claude Cowork. If you’re already familiar with coding agents such as Claude Code, you should try to use Cowork in the exact same manner. Essentially, focusing on isolating tasks, providing clear prompts, working in plan mode, and making sure the agent reviews its own output.
However, if you’re not familiar with coding engines from before, I’ll go through each of these points in more detail below.
Isolating tasks
Claude recently upgraded its most powerful model, Opus 4.6, to have a 1 million context window. A 1 million token context window is enormous, and you can fit a lot of context into the model at the same time. However, even though the model can’t fit 1 million tokens at the same time, I urge you to still minimize the token usage. The reason for that is that performance degrades the more tokens the model has in memory. In short:
The less tokens you have in the context, the better performance will be.
Now, you naturally have to balance this with the fact that not providing enough context to the coding agents will also degrade performance. For example, if you’re not providing information to the coding agent on which tools are available or details on how to complete tasks, obviously, the model will not perform well. The point I’m making above is that you should avoid unnecessary information in the context window of the model because that will degrade performance, not only because of the noise the model has to deal with, but also simply because the more text the model has to deal with, the more complicated the problem will become for the model.
An example of how to effectively reduce the size of the context window is to start a new thread or chat every time you start a new task. Simply put, a new task should never be done in the same thread as a previous task that you have completed.
It’s also good practice to reduce the amount of MCPs and connectors the model has available, as long as those MCPs and connectors are not relevant to the problem at hand. Anthropic has already dealt with this problem quite well by loading in such features dynamically only when relevant. However, simply removing unnecessary MCPs or connectors will eliminate the problem altogether.
If you look in the right side menu of Claude Cowork, you can also see a context tab highlighting everything that you have in context currently. This is shown in the image above. We can ensure that you remove any context that isn’t necessary anymore.
Lastly, another way to isolate tasks is that you always ensure you have different folders for the different areas you’re applying Claude Code to. For example, if you have a bunch of presentations generated with Claude Cowork, you should have a separate presentations folder that you always open whenever you want to make a new presentation. And if you have a separate sales project going on, you should also make a separate folder for that and work in that folder whenever you want to do sales with Claude Cowork.
Clear prompting
Clear prompting is another very important thing to keep in mind when working with coding agents. If you gave your prompt to a human and the human wasn’t able to understand the prompt and your intent with it, you’re making it hard for the model, which will again lead to worse results. Not providing explicit and clear prompts simply makes it less likely that the agent is able to implement what you intended.
Another important point here is to use plan mode whenever you start working on more complicated tasks. If you’re simply performing an easy task, you don’t necessarily need to use plan mode, but as a general rule of thumb, I urge you to always use plan mode whenever you start a new task. This makes the model think more about the problem you’ve given it, ensures that it asks clarifying questions so that you can elaborate on anything that is unclear, and, in general, improves the performance of the model.
You should imagine plan mode as a tool that aligns your intent with the agent’s implementation.
Reviewing agent output
Reviewing the agent’s output is not a very important thing to do if you want to get the most out of Claude Cowork. One of the great benefits of Claude Cowork is that it makes it easier to visualize things that are hard to visualize in a terminal where you use Claude Code. For example, you can ask the model to draw up illustrations or create diagrams. For example, if you’re working on a complicated architecture, you can ask the model to draw out the architecture, and it will create a nice flowchart for you.
The image below highlights how you can work with diagrams in Claude Cowork. In this instance, I asked Claude Cowork to generate a flowchart of a sales pipeline. And as you can see, the model creates the flowchart on the right-hand side, and I can continue working on the flowchart while having Claude Cowork on the left-hand side. This makes it incredibly easy to work with visuals inside Claude Cowork, which is one of the major advantages of working in Cowork rather than Claude Code.
Skills
The last part I want to cover is to use skills whenever you’re working on different tasks. If you’re performing any repeatable tasks, such as generating presentations, you should make it into a skill. This skill is essentially just a prompt that you have stored on how to perform a specific task. I, for example, have a presentation skill that knows exactly:
- What kind of theme do I want in a presentation
- Approximately the amount of text and text size I want in the presentation
- Where I typically want my images,
- The alignment I want in my presentations
And every other preference that I have.
Conclusion
In this article, I’ve gone into more detail on how to get the most out of Claude Cowork. Claude Cowork is a tool very similar to Claude Code that can essentially perform all the same tasks. However, it does have some benefits that Claude Code doesn’t have, especially for less technical users. I highlighted how you can get the most out of Claude Cowork by isolating tasks, using clear prompting techniques, and reviewing agents’ outputs. Even though I’m a technical user myself, I do find myself using Claude Cowork at times if I want to visualize results better or simply want to work on some tasks with a clearer interface. I urge you to try out Claude Cowork yourself to experience what it’s like, and you can determine for yourselves whether you want to perform some tasks in Claude Cowork or if you want to use Claude Cowork for everything. However, in general, I urge you to try to use coding engines to complete as many tasks as possible, as they are incredibly powerful models that are able to complete a lot of cognitive work.
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)
📚 Get my free Vision Language Models ebook
💻 My webinar on Vision Language Models
👉 Find me on socials:
💌 Substack