Introduction
A warm welcome to another episode of TJ's Technology Tuesday and Digital 4 Productivity, the podcast for productive digitalisation. Today I would like to share my experiences with three systems that I have worked with very, very intensively over the past few weeks. Because, as you know, what is not relevant to me is what someone claims to know about AI. And that is exactly how I feel again and again: the more I know, the more I have the feeling that I actually know nothing at all. In other words, I find it almost impossible to keep up.
Three AI tools, three different strengths
That makes it all the more important to focus on the things that really work and to implement them consistently. And that is exactly what I do on behalf of my clients: we look at how we can implement digitalisation productively. There are three tools I would like to say a little more about. The first tool is one I have rediscovered, namely Langdock. Spelled Langdock, a German company.
Langdock as a bridge between AI technology and productive use in the company
The idea behind Langdock is that I have a platform there where I have all the common AI systems, that is OpenAI, ChatGPT, Mistral. I have Claude, Anthropic, I have Grok there, all the various models – but all with a layer in between, so that the data does not leave Germany, or rather so that I can use it in a GDPR-compliant operation. In the past, Langdock was always a bit like other systems for me, somewhat limited in performance compared to full access. And that is why I was initially more drawn to OpenAI, then of course very strongly – and I still am – to Copilot, and thirdly to Claude Anthropic.
Is Langdock GDPR-compliant?
Of course there are sometimes things – I recently had something highly confidential for a client again – where I keep saying: okay, in that case it is of course quite good to either work with Copilot, where I can say I stay with the data in Germany. The data is already inside the Microsoft 365 universe anyway. That means I do not create an additional separate pool of data. That is one starting point. Of course the AI processing takes place in Ireland, which is again out of Germany, but still within Europe.
And I always have the risk that I am working with American companies, with the potential reach of the USA under the Patriot Act. Those are of course the risks.
Langdock – an extremely good AI tool
So, secondly, and this is something where Langdock surprised me very positively. In terms of quality, Langdock has become extremely good by now, because, similar to Copilot, it is optimised for business use. And when it comes, for example, to me saying I want a certain source format of information and want a certain result format on the other side, I was able to build myself an agent there – similar to the Copilot Agents or to GPTs – where I gave it a few patterns, a few results that were produced from these patterns, and said: based on this pattern, please pull information out of these documents and prepare it for me accordingly.
I let several AIs run in parallel and Langdock was by far the AI that did it best, both in terms of the result and the formatting. Because what I often experience with Copilot, for example, is that I do get a good result, but the formatting still requires a lot more manual reworking, that I cannot output it as Word. And here I get it across right in the first step as Word. Langdock also costs, I believe, 25 euros a month. So in terms of licence costs, given that I even have several models there, it is very, very good.
In terms of the result, I even liked it a little better than Nelly AI, which has a similar approach. What I would like to pass on to you in the spirit of sharing experiences is: do try out different systems as well. It also has the additional charm that, if you regularly hit the caps somewhere – as happens to me, the capacity limits – then you can carry on with another AI tool. And Langdock – I do not get any commission for this, I am not a Langdock partner either, but I really do find it interesting and worth a test. And, by the way, if needed you can even connect it to your SharePoint, your Outlook and all these things.
Naturally, I once again have a gateway here with corresponding risks. But you have the option to connect such things and, against that backdrop, to work productively. But even within Langdock alone, I have to say, I liked it very, very much. Which to some extent once again underlines for me that it is a good idea to take another look at tools that you may have set aside in their first or second version and to say: hey, maybe they have evolved in a positive direction. Just like Copilot has just been released in a new version.
I will test that a bit more this week. I would be happy to give you another update on that next week. Perhaps you have already seen it. It already simply looks a little different. It is also supposed to deliver better results in the Office applications.
Let's see how good the results are. I will report on it next week. So Langdock, that is the first thing I would like to pass on to you as a piece of experience.
Now, the second insight I would like to share with you is my by now quite intensive work on the topic of vibe coding, that is, developing programs yourself. I work with Lovable a great deal here, and here too I have had a very positive experience.
Langdock versus Lovable versus Claude
At the beginning with Lovable I always had the challenge that it made a great many mistakes. It dutifully corrected them again, but there was a mistake in almost every screen. And now it is the case that mistakes still occur from time to time, but significantly fewer. So you can tell, on the one hand, that they always keep pace with the current model from Claude, that is 4.7, and now 4.8 is the new model. That means the Claude model from Anthropic keeps getting better.
Secondly, presumably their implementation as well. And the nice thing here is that you can always prompt on the left-hand side what you want. On the right-hand side you see what the result of the program looks like. And what I did for a client, for example, is that I built a program, an individual program for him, to the point where it was roughly 95% finished, roundabout. In other words, an administration system for training courses, for occupational safety training.
So I built that to about 95%, 99%, let's say 95% finished. Then I held a Teams call with the responsible training manager and walked her through it, showed it to her, we recorded it so she could look at it again. And then we said: now there are three options. Either you now have some small thing – the button should be green or red, should go from left to right, should be bigger, should be smaller. Then I said: just prompt in the left-hand line and say, make the button different. The client can do that themselves up to a certain degree.
But please not from scratch, because I am also no fan of starting in Lovable; instead, please always start in Claude, because Claude is optimised for programming tasks. And having a very clean, dedicated structure there naturally helps. After all, I have introduced, let's say, three CRM systems, among others at Tchibo Kaffeeservice and back in the day at Vitality. I have carried out SAP migrations. So against that backdrop I have a great deal of experience.
What should you bear in mind when programming with AI?
At 14, back when I got my C64, I programmed on that machine, was already selling programs and so on. So that is why I have a lot of my own programming experience; I am out of practice because I no longer program today, but the structured thinking helps me and I know which concepts matter. So, that is why it is important: start in Claude, then move over to Lovable. You should have someone who has programmed before. That is also why I do not believe that programmers will be out of work; rather, programmers are the perfect vibe coders, because they know exactly what data tables are, what databases are, which ones to use, how to set up certain things cleanly in terms of security, what a firewall is, so to speak how to harden certain things, what encryption is and such things.
Users often do not know that. And that is why you then think of certain things. Lovable has automatic security checks, which is already good. But often even I do not understand the questions. The good thing is that you can always switch back over to Claude and then ask Claude: tell me, explain to me what I should do.
Here it is important, as is the case so often with AI, not just to say: AI, get on with it, but rather: AI, explain it to me in a dialogue. So once again I have learned very, very much about current programming, about current security technology and so on. And the cool thing is, I said the first thing is that you can do it yourself. The second issue is, you are not sure whether something will work? Then there is a so-called Plan Mode in Lovable.
That means it does not get going straight away; instead, as with many AIs, you can also say, let's plan first, let's discuss first. And once it arrives at a solution where you say, oh, that's great, go ahead, implement it, then you, dear client, can do that too. And if it is then the case that you say, oh, that is somehow too complex, or I do not understand the question, or the answer that comes up in Plan Mode, then the idea is that we have a Planner board in Microsoft Teams and then we say: okay, new requirement, I receive it, then I move it to in progress, then I say ready for release, it is released and then the matter is done. So the nice thing is that with vibe coding you can develop software together with people who understand development. Because I am not a great fan of putting a pure developer on it; they are too far removed from the user.
I am also no fan at all of users who have never programmed simply vibe coding away. So at the beginning I, too, made the mistake of starting in Lovable, and completely different results come out of that. Quite apart from the fact that you burn too many tokens. So you end up getting a workflow out of it; and if you work out program briefings with Claude more often, then Claude of course also learns in this area. So, excellent.
My recommendations for using Langdock and Lovable
Yes, so the first is Langdock, the second is Lovable, that is for the topic of individual programming. The third is my new friend that I have here on my phone. My new friend on the phone is Hermes. Hermes, which I have integrated there in the Telegram bot. Perhaps you remember from my episodes on the topic of OpenClaw – it used to be called ClaudeBot, so it is called Moltbot – this system invented by Peter Steinberger, a bot that also does things autonomously.
And here there is a kind of further development; it is a different development team, from the USA, who made it. They developed Hermes and the idea of Hermes is basically the same. So it is also an assistant that either runs on a local machine or – which I always recommend – I have the thing installed at Hostinger. So it runs on a separate German server in Germany in a Docker container. And here, too, once again: if you do not know what a Docker container is, what a Virtual Private Server is, then you should keep your hands off it.
Please also do not install it on your own computer. But if you know a little about what you are doing, then it is relatively easy to set up. I find it a little easier to set up than the Moltbot. And it has one huge advantage. The huge advantage is that it has Learning by Design built in.
That means, when it completes a task, it automatically always runs a feedback loop. In other words: how can I improve here once more? And it updates its own skills, that is, these capabilities. Beyond that, you can of course also say: do you notice that? Are you saving that?
And it does that extremely well, I have to say. Significantly better than the Moltbot. I have tested both, after all. As you know, I also do not let a non-swimmer explain to me how to swim faster. So I always test these things myself first.
How do I use Hermes AI?
And Hermes is really in productive use with me. Perhaps you already saw the first LinkedIn posts from me on Sunday and Monday, where I made these LinkedIn carousel posts. I never used to make those, because I would say: ah, generating some 7-page PDF is always work and fiddly. Today Hermes does that for me. In the morning it makes me a morning briefing, where I get the most important news, where I get a current AI study with links every day.
And then, in the morning, alongside my morning coffee, I can already tap on it somewhere and read the most important news, thereby staying up to date, reading a fresh new AI study on the topic every day. And then, out of this news, it also makes me a 7-page PDF. By the way, on the last page there is not only a photo of me, not only the contact details for follow-up questions, but it also says "Assisted by Hermes AI, trained by Thorsten Jekel." You should get into the habit of this when you use AI systems for support: please always label it as well. A colleague from the speaker industry recently commented on my comment, on my post, where I once presented this and said: how do you label this?
The comment came: but you don't even have to label it, so to speak. Yes, maybe I do not have to label it, but I think it is ethically appropriate that you label it when you do things with AI support. And it also makes a difference to me whether I say it is simply AI-Grok or whether it is, just like this Hermes bot. Until I had it to the point where I need it now, it took me about three weeks, where every day I sharpened it again, where I said: please look at the sources, please not just Google News, but look at Stanford, look at Harvard, look at MIT, look at this and that.
So we have, I would say, those 14 days, three weeks, where we corresponded very intensively with each other, hence trained. In the same way – and I will also show this with Claude Design in one of the next episodes – I can now also create offers in such a way that, on the basis of a Teams transcript or a podcast transcript, I can simply quickly say: bang, make me an offer. But all of this with intensive training. That, too, is to some extent the bracket of today's episode. The important thing is: do not use too many tools, also revisit the tools you put aside earlier, but then also say that ideally you should have one tool per task.
Conclusion
The important thing in all of this: first switch on your brain, then the technology. So first work conceptually, then try the things out, then also label them. Yes, next week I will say something a bit more in depth about the topic of the new Copilot here. It keeps evolving too. And should you need support from me as a Personal IT Coach for executives, for keynotes, for your staff, for your clients, for your partners.
Be it C-level sparring, where together – just as I am currently doing for a subsidiary of Würth, for example – with the owner, with the managing director, we simply say here: we are driving the IT systems forward here and are slowly overtaking the Würth colleagues. They are getting really curious about what great things we are doing here in our little Gallic village. You simply pick up opportunities along the way. Or should you, as I will be doing again shortly, be at the Volks- und Raiffeisenbank, where I can train Microsoft 365 there in such a way that you really use it productively, then you are welcome to receive my support. In this spirit, I wish you a productive Tuesday and a productive week.
First switch on your brain, then the technology.
Yours, Thorsten Jekel.
Key Takeaways
- Langdock is a GDPR-compliant AI platform from Germany that bundles several models (OpenAI, Claude, Mistral, among others) with a data-protection layer and costs around 25 euros per month.
- Langdock allows you to build your own agents with sample inputs and delivers results directly as a Word document – which significantly reduces reworking compared to Copilot.
- Lovable is suitable for vibe coding (AI-supported software development), but should always begin with a Claude planning phase rather than being started directly in Lovable.
- Programming experience is a decisive advantage in vibe coding: those who know concepts such as databases, security and encryption achieve significantly better results.
- Lovable's Plan Mode makes it possible to plan and discuss changes first before they are implemented – ideal for collaborating with clients.
- Hermes is an AI assistant in a Telegram bot that runs on a VPS (e.g. at Hostinger) in a Docker container and has built-in learning feedback.
- Hermes automatically creates daily morning briefings with current AI studies as well as LinkedIn carousel PDFs – after around three weeks of intensive training.
- AI support should always be labelled transparently, for example with the note "Assisted by Hermes AI, trained by Thorsten Jekel."
- The recommendation is: do not use too many tools at once, retest tools you previously set aside, and ideally use one dedicated tool for each task.
- The basic principle for all AI applications: first think conceptually (switch on your brain), then deploy the technology.
Frequently Asked Questions
What is Langdock and who is it suitable for?
Langdock is a German AI platform that bundles several language models such as OpenAI, Claude and Mistral under a GDPR-compliant data-protection layer, so that data does not leave Germany. It is particularly suitable for companies that want to use various AI models in a data-protection-compliant way in their business operations and to build their own agents.
What does Langdock cost and which models are included?
Langdock costs around 25 euros per month and offers access to several common AI models, including OpenAI, ChatGPT, Claude from Anthropic, Mistral and Grok. The price is rated as very attractive, since several models are included in a single package.
How does Langdock differ from Microsoft Copilot?
Langdock delivers results directly in Word format and performs better than Copilot both in terms of the result and the formatting, where manual reworking is often required. Copilot, on the other hand, has the advantage that the data is already located in the Microsoft 365 ecosystem and no additional pool of data is created.
What does vibe coding mean and how does it work with Lovable?
Vibe coding refers to the AI-supported, self-driven development of software, where you enter requirements on the left-hand side in Lovable and see the running program on the right-hand side. For the best results it is advisable to first work out a clean program briefing in Claude and then implement it in Lovable.
Do you need programming knowledge for vibe coding with Lovable?
Programming knowledge is a clear advantage in vibe coding, because you know concepts such as databases, security, encryption and firewalls and can deploy them deliberately. Users without any programming experience should not start directly in Lovable, as important security-relevant aspects can otherwise be overlooked.
What is Plan Mode in Lovable?
Plan Mode in Lovable ensures that the tool does not begin implementing straight away, but first creates a plan and puts it up for discussion. Only after approval by the user are changes implemented – which is particularly helpful when collaborating with clients and with complex requirements.
What is Hermes and how does this AI assistant work?
Hermes is an AI assistant that is integrated as a Telegram bot and runs on a Virtual Private Server in a Docker container. It has a built-in learning feedback loop that, after every completed task, automatically checks how it can further improve and updates its own capabilities.
What concrete tasks does Hermes take on in everyday work?
Hermes creates daily morning briefings with the most important news and a current AI study including links, as well as automatically generated seven-page LinkedIn carousel PDFs. Until this bot reached the desired quality, around three weeks of intensive training were necessary, during which sources such as Stanford, Harvard and MIT were deliberately set.
Why should AI support in content always be labelled?
Labelling AI support is important for ethical reasons and signals transparency towards the audience. Beyond that, it makes a difference whether a piece of content was simply generated by an AI or whether a trained, individually configured bot stands behind it.
What is the basic rule when using AI tools?
The central recommendation is: "First switch on your brain, then the technology." This means working conceptually first, defining the right requirements and then deliberately deploying a suitable tool – instead of starting directly with the tool and hoping for good results.
Tools & Resources Mentioned
- Langdock – GDPR-compliant multi-model AI platform from Germany, approx. 25 euros/month
- Lovable – vibe-coding tool for AI-supported software development
- Claude from Anthropic – the AI model underlying Lovable, optimised for programming tasks
- Microsoft 365 / Copilot – as an alternative for data-protection-compliant AI use in the Microsoft ecosystem
- Hermes – AI assistant as a Telegram bot (self-hosted on a Hostinger VPS, Docker container)
- Moltbot / OpenClaw – predecessor assistant by Peter Steinberger, against which Hermes is compared
