Introduction
Welcome to another episode of TJ's Technology Tuesday. Today, once again live from the studio in Berlin, having just stepped out of a Power Automate training. I assume you are watching too, dear Christian. Good to have you back. When I glance to the left, I of course always keep an eye on the comments as well.
What are we covering today? Last week we looked at the whole topic of automation and AI agents. Today I would like to discuss the next step with you: how can you program with AI? But before I dive into programming, let me offer a correction and addition from last time.
Addition to the last podcast episode:
Facebook's acquisition of Manus was rejected
Last time I told you that, in my view, Manus as an agent is not usable — or only very limited. The reason: Manus was acquired by Facebook. Manus was originally a company based in Singapore, and now it is back to its roots. Chinese regulators have blocked and rejected Facebook's acquisition of Manus. As a result, that deal now has to be unwound.
This means Manus will soon no longer belong to Meta. In that sense it is once again Singaporean-based, but at least it is no longer part of the Meta ecosystem. And that is always an important topic — regardless of which IT tool or AI tool you use, one of the essential questions for me is always: who is behind it? You will recall what I said last week about Anthropic and Amodei. The backstory is that the founder of Anthropic, Dario Amodei, and his sister were previously at OpenAI, the maker of ChatGPT. They went independent with the goal of building a more ethically sound AI.
What does that mean in practice? It means that with AI today, what the AI can do is almost less important than the guardrails — the boundaries on the left and right that constrain it. Experts broadly agree that setting certain limits is a pretty good idea. Just think about what happens if someone asks: How do I build a bomb?
It is a good idea for the AI not to answer that. Or — I don't know whether you caught the recent Meckel and Mattes episode, which is very much worth listening to — they discussed the topic of youth protection in the context of AI. The clear argument there was that an AI should under no circumstances encourage suicide, particularly for young people who may be dealing with emotional difficulties. That is a classic example of guardrails. There are different approaches to implementing them.
If you look at Grok, Elon Musk's AI, the philosophy is no guardrails — as much freedom as possible. Free speech, taken somewhat to the extreme. If you look at OpenAI, there was a major conflict back in the day — the reason why Amodei went independent — not only because a non-profit organisation was converted into a for-profit one, but also because the goal of building a safe AI was eventually thrown overboard in that process. As a result, Amodei went on to found Claude. That is why I am a big fan of Claude. And this brings us to the topic of programming, because I have introduced Claude multiple times as a system that offers you three main options — and beyond those, even more.
Options for programming with Claude Code
Let us look at the three main options Claude offers. First, there is the pure chat — similar to what you know from OpenAI and ChatGPT. You ask a question, you get an answer. Then there is the Cowork area, for which you need a paid subscription.
That starts at 15 euros per month and you need to install the app, which is available for both Windows and Mac. You can then say things like: sort my emails, or organise files in a folder, rename them, perform specific tasks. You can say: translate a particular presentation, for example. And then there is the third area, called Claude Code. This can go even deeper, issuing commands at the terminal level on your system.
The idea is that you give Claude Code access to local folders and files. It can only operate within that folder and will always ask before performing anything critical. You can configure this. This is where it differs somewhat from tools like OpenClaw, Moltbot, or whatever they are called — the kind of autonomous bot you may have heard of.
What are the limitations of AI agents compared to Claude Code, Moltbot & Co.?
The argument was: do I even need automation tools anymore, do I need AI agents at all — can't Moltbot just do everything on its own? Well, the thing is: you want to keep a completely autonomous system like that as far away from your productive systems as possible. Do you really want it to decide: those emails are no longer needed, let me delete them? Do you want it to gain access to your email system without your approval? Those are the kinds of situations I would call highly dangerous.
Anthropic's Mythos model can uncover significant security vulnerabilities
In this area you really need to maintain a good balance between productive use of systems and security. And as you may have noticed, Anthropic's new model called Mythos has not yet been released publicly, because this model is capable of very rapidly uncovering significant security vulnerabilities in core systems — including very old ones. Anthropic first made the model available to those who are affected by these security issues, so they can fix them. So Anthropic are no saints either — they want to make money, and there are critical voices about that too — but they do at least have an ethical foundation that, in my view, is better than what their colleagues at OpenAI are doing. What is also interesting about Claude is the ability to use it increasingly within Microsoft as well.
What the paid Claude subscription offers
You can now optionally use Claude models in Microsoft as well. With a paid Claude subscription you can use plugins in Excel, in PowerPoint and now even in Word. These are excellent — genuinely good, and significantly better than the ones from Microsoft. And Cowork — the mode where you instruct it to handle specific tasks — will now roll out in Germany in stages as part of an extended E7 licence. This means you can automate certain tasks.
A practical example of programming with Lovable
Now, many things can also be automated with Power Automate, and I want to illustrate the topic of programming with a very concrete example. Because I believe that, rather than tinkering endlessly with complex solutions, you can now build them quickly using low-code and no-code tools. One system I highly appreciate is Lovable. Lovable is a Swedish company. Let me switch to demo mode now before I open up the live view.
Here we are. Let me switch over to the Lovable view. What is the idea behind Lovable? The idea is that you have two workspaces — let me go to the right-hand side. On the left you have the workspace where you type what you want it to do.
And on the right you have the preview area showing the result. What you see here is a fairly complex application, but let me navigate to a different project — something much simpler that gives you an idea of what these tools can do today. Let me go to the dashboard and open a very simple application I built during a training session for the Schmidt Entrepreneurship College. I simply said: I want to build a web app. Specifically, a company analysis tool. I'll make the view a bit larger so it is easier to see. The goal was a web app with a questionnaire based on a specific framework — I simply photographed that questionnaire. You can even see a slight shadow from the phone. Then I described what the output should look like. One prompt, and that was the result.
So here is the company analysis page. I clicked Start Analysis and now I can use these sliders to adjust various parameters, choose which direction to go, do this for different areas and click Show Evaluation. I then receive an evaluation — and I can even export it as a PDF.
That was one prompt. I did have to make one small correction — there was an issue where the fill colour was not filled in. One correction. That was it.
Now, this is a relatively simple example, but you can already see that in the past you would have needed a large piece of software to achieve this. Things are clearly much easier now. And if I look at my dashboard and consider what I have developed for myself over the last four weeks, it is a CRM system.
So — what did I do with this CRM system? Let me just log in for a moment — bear with me, back in a second.
Let me go into demo mode. There we go. And this is also a nice feature. I simply told it: create a demo mode — so anyone can see it. I have built quite a comprehensive CRM system here with activities, with companies, with contacts for those companies, and I can send emails, export things as PDF, enrich data to pull in additional information.
If I have LinkedIn information, I have a LinkedIn tab where I can even access LinkedIn messages and similar things. I have sales opportunities showing open proposals and my pipeline. All demo data, currently — I am in demo mode. I have events showing people registered for certain events. I have a calendar with calendar integration, and I can even build automations into it. It is quite a capable CRM system, with integrations for Microsoft 365, LinkedIn and WhatsApp for Business — plus webhooks. Not entirely uncomplicated.
What approach do I recommend when programming with AI?
What is the most sensible approach? I did not start by saying: just build me a CRM. What I did first was go into Claude and start with an idea — and this is a good approach in general: kick off with an idea.
What I did was look at a CRM system that was close to my requirements. Then — by the way, if you can no longer hear anything, the audio should appear if you click the speaker icon in the bottom right. That should work by clicking — I also included this in the notes. Dear Barbara Zimmermann, I hope this is visible; otherwise we will correct it in the recording.
So I went to a website and looked at a CRM that was relatively close to what I wanted. That was Monday CRM, which I found quite appealing.
I then went through Claude and said: create a briefing for me based on this. I went through the individual sections and said: draft me a briefing. It produced a first version — already a very good document, because I would certainly not have been able to describe all this as well as the tool does. So it created quite a detailed document from the start.
But then I said: I want to go further. I want certain things removed that I do not need, and certain things added that I do want. This went through several iteration stages. Let me fast-forward — until eventually a document at version 8 emerged. I also had names suggested afterwards, which I reviewed. But just looking at the end of that process, I had version 8 here — the final briefing. And at that point I had a very, very solid briefing.
This means I ended up with a proper requirements document — a functional specification — with clean technical implementation details and everything that goes with it. What I did next was go into Lovable, and right at the top, going back to the very beginning, I said: take this document and build a CRM on the basis of it. Use the five-stage process model I included as a foundation. And then I just kept working my way through step by step. The idea is that I can always flag a bug, correct it with a prompt, and work through iterative loops repeatedly.
And that is really the beauty of it — you can work through things incrementally and never have to worry about how to actually implement something. The great thing is that Lovable uses Claude as its AI engine. So it handles briefings that were created with Claude extremely well.
Another interesting aspect is that I can access the entire code here. So if I know what I am doing, I can edit this code directly. And what is also interesting is that I can attach a backend to it — with cloud infrastructure, the relevant tables, databases and so on. And I always have the option of going back to the preview, synchronising the whole thing with my GitHub. GitHub is a system where you can sync and save the code from Claude. This gives me the complete code and I can revise it. Interestingly, I can now also let Claude Code optimise it further — or I can have it revised by a human developer.
Six months ago I would have said: if you want to scale something like this broadly, always have a developer do it. I would never have done that without a programmer. But things have changed. Anthropic itself, for example, no longer programs anything manually.
They have built their software in such a way that they can continue developing their AI with it — and the fact that we can use it for our own purposes is essentially a by-product. Of course, the right thing to do is always to have an experienced developer take a look.
If you are using the application for yourself, then honestly, you no longer need traditional programming at all. If you are selling it, I would still have it checked more thoroughly, because in that case you carry a higher risk of causing harm to customers — up to and including financial damages for which you could be held liable. So that is certainly something to be careful about. But we have reached a point in programming where the code being generated is not only not worse — in many areas it is actually better. And particularly when it comes to security, asking the model how to improve security: in that domain today's models outperform humans.
Conclusion
To everyone among you who is a programmer: I believe that the relevant competence today is less about knowing exactly how to write syntax and all those details. What matters is being able to create a very good briefing and to evaluate the output. A regular user cannot answer the questions Claude asks, let alone judge the code it produces. So I do not believe programmers will become obsolete because of AI. I do believe, however, that programmers who use AI — to avoid getting bogged down in trivial details — will be significantly more productive and will outperform programmers who do not use AI.
This is a pattern we see in many fields — in the legal profession, for example. Volker Römermann recently wrote an excellent book on the subject. The line goes: AI will certainly not replace a lawyer, but a lawyer who uses AI will replace a lawyer who does not. Not today, but in the near future.
With that in mind, I do not want to cause fear, but to encourage you to use AI as a smart tool. Always with the reminder: engage your brain first, then the technology. Your personality coach for executives, Thorsten Jekel.
Key Takeaways
- AI-assisted programming enables even non-programmers to create complete web apps and CRM systems using natural-language prompts — without deep coding knowledge.
- Claude offers three main modes: Chat, Cowork (from €15/month with a desktop app) and Claude Code for deep terminal-level access to local files.
- The tool Lovable (a Swedish company) lets you develop full web applications through natural-language prompts and refine them iteratively.
- Recommended approach: first create a detailed briefing in Claude (ideally through several iteration stages), then use that briefing as the foundation in Lovable.
- Lovable uses Claude as its AI engine — which means it handles briefings created with Claude particularly well.
- Generated code can be synchronised via GitHub and subsequently optimised further with Claude Code or by a human developer.
- For private or internal use, AI-generated code is often sufficient; for commercial products, an additional review by experienced developers is advisable.
- Autonomous AI bots (e.g. Moltbot) should be kept away from productive systems, as they can perform critical actions without approval.
- Anthropic's Mythos model has not yet been released publicly because it is capable of rapidly uncovering significant security vulnerabilities in core systems.
- AI will not replace programmers — but programmers who use AI will be significantly more productive than those who do not.
Frequently Asked Questions
How can you program with AI without knowing how to code?
With tools like Lovable you can create complete web apps by explaining in natural language what the application should do. A single prompt can already produce a working application, which is then improved iteratively through further prompts.
What is Lovable and how does it work?
Lovable is a Swedish no-code/low-code tool for developing web applications that uses Claude as its AI engine. You describe what the app should do in the chat window and see the result immediately in the preview area — errors can be corrected with a prompt.
What are the three main options Claude offers?
Claude offers: first, pure chat (similar to ChatGPT); second, Cowork mode (paid subscription from €15/month, with a desktop app for Windows and Mac, for automated tasks such as email sorting or file renaming); and third, Claude Code, which can issue deep commands at the terminal level on your local system.
What is the recommended approach for AI-assisted programming?
The recommendation is to first create a detailed briefing in Claude — ideally through several iteration stages. This briefing is then used as the foundation for the actual development tool (e.g. Lovable) to build the desired application step by step.
Why should autonomous AI bots be kept away from productive systems?
Autonomous AI bots like Moltbot can perform critical actions without explicit approval — for example deleting emails or gaining access to systems. It is therefore strongly advisable to keep such systems strictly separate from productive environments and always build in an approval step.
What distinguishes Anthropic and Claude from other AI providers such as OpenAI?
Anthropic was founded by Dario Amodei and his sister, who previously worked at OpenAI and left to build a more ethically sound AI. While OpenAI converted from a non-profit to a for-profit organisation and reduced its safety focus, Anthropic places greater emphasis on ethical guardrails.
What is Anthropic's Mythos model and why has it not been released yet?
Anthropic's Mythos model has not yet been publicly released because it is capable of very rapidly uncovering significant security vulnerabilities in core systems — including very old ones. Anthropic first made the model available only to those affected by these security issues, so they could close the gaps.
Can I use Claude plugins in Microsoft Office?
Yes, with a paid Claude subscription you can use Claude models as plugins in Excel, PowerPoint and Word. According to the presenter, these plugins are significantly better than the AI features offered by Microsoft itself.
Will programmers become obsolete because of AI?
Programmers will not become obsolete because of AI — professional expertise is still needed to create a good briefing, evaluate the generated code and answer the AI's questions meaningfully. However, programmers who use AI will be significantly more productive than those who do not.
When should AI-generated code be reviewed by a human developer?
For private or internal use, AI-generated code is generally sufficient. However, anyone planning to sell the application commercially or pass it on to customers should have it reviewed by an experienced developer — because errors can lead to financial damages for which you may be held liable.
Tools & Resources Mentioned
- Lovable – Swedish no-code/low-code tool for creating web applications via prompt; uses Claude as its AI engine
- Claude (Anthropic) – AI platform with Chat, Cowork and Claude Code modes; from €15/month for the paid subscription
- Microsoft 365 – Claude plugins available in Excel, PowerPoint and Word
- GitHub – version control system for synchronising and saving code generated with Lovable
- Monday CRM – used as a reference CRM to create the briefing for the self-built CRM system
