How I use a coding app (Claude Code) for my daily knowledge work.

Friday, I had Claude fetch the configurations of 10 custom GPTs. A task that would have taken me more than 2 hours manually was done in 30 minutes by the AI. The time saved, however, was not the big surprise. The big surprise was that it was actually able to go on the internet and complete this task, working on its own for the 30 minutes, and completing an actual task for me.

That experience made something very clear to me: the difference between these tools isn't the interface or the features. It's how much work you're willing to let happen without you.

The day before, I sat down with Greg and Arne, both partners at Generative AI Strategy. The topic of the meeting; Why should we use Claude Code (terminal), Claude Code (app), or Claude Cowork if we are not writing code.

In this week’s newsletter, I try to answer that. It is not a summary from our conversation, but instead aims to be a structured representation of my thoughts after it.

1. The history behind Claude Code (Terminal), Claude Code (App), Cowork, and the Chat interface

By now, anybody that is reading this newsletter will be used to the chat interface. We know this from Chat GPT, Gemini, and also of course Claude.

These tools are valuable because they understand and create. You give them information, they give you something new.

Then, at the end of last year, Claude Opus 4.5 dropped within Claude Code. With it: a new form of AI. Agentic AI.

We already mentioned that traditional Generative AI tools interpret and create, agentic AI tools are different. These tools don’t just interpret and create. They interact with the real world.

At first, the only way to access Claude Code was through the terminal. Then, on November 24th 2025, Anthropic released the desktop version alongside Opus 4.5. The main difference? The interface.

The terminal offered a hacker-like tool, available to programmers and real early adopters, many people felt overwhelmed with this interface. But, as the tool was integrated in the desktop app of Claude, it saw its real liftoff.

Suddenly, developers were not writing code anymore, they started to fully trust the AI to do it.

But people started using it more and more for other use cases. As the creator of Claude Code mentions:

“we saw people using it for all sorts of non-coding work: doing vacation research, building slide decks, cleaning up your email, cancelling subscriptions, recovering wedding photos from a hard drive, monitoring plant growth, controlling your oven.”

Therefore, on January 12th of this year, they released Claude Cowork, which is the “Claude Code for the rest of your work.”

Three tools, one model. The real difference? How long you're willing to look away.

2. Understanding the differences between the tools

Now we understand the history, we can work out the differences between the tools.

The obvious difference is that chat interprets and creates, while the other three interact with the world. But the more useful way to think about it: these are points on a spectrum of delegation. Chat is 'help me think.' Cowork is 'do this task.' A configured terminal setup is 'handle this while I'm gone.

As for the differences between the other three, however, the most noticeable difference is the interface. With the terminal version of Claude Code being the most raw interface, while the interface of Cowork is the most simplified one.

Another important difference is the safety of the tool. Like I wrote in an earlier newsletter, giving an agentic model access to your laptop is like giving a guest unlimited access to your office. It can delete any file and move any folder it wants. Because of the grave dangers associated with this, Anthropic decided to add deletion protection into Cowork.

3. Understanding the role MCP’s and skills play in an agentic world

Interface and safety aren’t the only differences. There’s also customisation.

The desktop apps come with many capabilities built in: browser access (which enabled the GPT configuration pull I mentioned at the start), PowerPoint generation, and Excel files.

In the terminal, you can also get those capabilities, but they are not always included. You have to add them yourself, by giving your Claude Code access to MCPs, and skills:

• MCP (Model Context Protocol): Think of it as a universal adapter plug. It lets Claude connect to any app or service (like Notion, Slack, or Outlook), giving it access to your personal context without you having to copy and paste it in.
• Skills: These are pre-built recipes you add to Claude. This can include skills for how to write emails in your personal style, but also for how to create meetings or write the storyline of your next presentation.

To give you an idea: the standard Claude Code can write a presentation. My customised version pulls my last three client decks from Sharepoint, checks Fireflies for what was actually said in the kickoff call and emails, and writes in my style. Same model. Very different result.

4. Choosing between the three by deciding how much to delegate

We’ve covered the model, interface, safety, and customisation. Now: how to actually choose.

Choosing between the three isn't really about features. It's about asking yourself: how much am I willing to let happen without me? The answer to that question picks your tool. Just like you ask different colleagues to help you on different tasks, different agentic tools are good at different things.

But, to help you get some ideas and make a decision, I would like to end off with how I personally use the different tools.

I. Claude Code terminal

I use this for my most customised workflows:

1. Chief-of-staff: This agent is connected to Outlook, Fireflies, and Notion through MCP (each of our employees has a different setup). It helps me organise my week. I can ask it to prepare my meetings, check which emails I have to follow-up on, and which tasks are left on my to-do list.
2. Slide builder: This is one of the agents that saves me most time during the week. I gave it the skill to write storylines, with many of our existing presentations in its context. It can also research content for additional slides. And then create a layout based on all of it. Lastly, it is connected to the Nano Banana MCP, which enables it to create stunning images for the slides it helps me create

II. Claude Code Desktop

Honestly? I rarely use this one. The terminal covers my customised workflows, and Cowork handles everything that needs standard tools. Desktop sits in the middle. More accessible than terminal, more flexible than Cowork, but I haven’t found a task where that middle ground is what I need.

That said, if you’re comfortable with some technical setup but don’t want the raw terminal experience, this might be your sweet spot.

III. Claude Cowork

This brings me back to where I started: fetching those GPT configurations. Cowork shines for tasks that need browser access and standard tools. No custom setup required.

I’ll be honest: I haven’t tested this one as extensively as the terminal. But for anything involving the internet, research, data extraction, filling out forms, pulling information from web apps, this is where I’d start. The deletion protection also makes it safer to experiment with.

More to explore here. Which brings me to the last question from our conversation.

The things we don’t know we don’t know

The real unlock isn't picking the 'best' tool. It's recognising that you are the bottleneck, and deciding how much you're ready to let go.

My goal for the coming weeks: extend the leash. Longer tasks, less supervision. A finished presentation while I take a walk. Possibly with the help of tools like Openclaw (more on that in my weekly substack).