Skills Not Hacks!

Transcript

00:00:00Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:00:07Two technology-loving minds who not only talk about artificial intelligence but live it.

00:00:14Here there are clear classifications, real practical insights, and a fresh perspective on what is possible.

00:00:20Understandable, critical, and always with a wink.

00:00:24Food for thought, to chuckle, and above all to discuss.

00:00:33Welcome to a new episode of Singtiff and Sing AI.

00:00:37We thought today we would talk about how to use skills to prompt less.

00:00:44Today I'm once again not alone, but today I'm in familiar company, namely without a guest,

00:00:52but instead with Jens. Hi Jens, great to have you here.

00:00:56Hi Mark, yes, every now and then it still happens that I can't be in an episode.

00:01:02I'm already looking forward to the episode when I will be here alone and do an episode.

00:01:07You could be very curious about that.

00:01:09I don't know the topic yet, but I'll think of something.

00:01:11This time I'm really glad to be here, Mark.

00:01:14And we really have such an exciting topic,

00:01:17because we've been talking about new bot forms and such things lately.

00:01:22We're talking about what connects all of this, no matter which company these things come from,

00:01:27that now agent skills are somehow different, I've said that before, this word skills,

00:01:30you just mentioned it too, today we want to shed some light on it. What are

00:01:34these skills, these abilities that I can give to my agents so that they in the

00:01:41end deliver better results? You've already said something important, it's

00:01:46not about the skill I have as a person, that I’ve gotten some certificates

00:01:52from somewhere and can prove that I've attended great trainings, but rather about the possibility

00:01:57to give skills to the systems so that we can move away from that. I mean, what was the

00:02:03term that was all the rage a while back, yes prompting strategies, prompt frameworks,

00:02:10prompt engineering. At the end of the day, all of that ends up being reflected here, but I would

00:02:16suggest we start with a little overview. What are skills actually,

00:02:22I'm going to say now, include, before we might take a deeper look.

00:02:27I will put something in the room and then you can correct again. So I would say,

00:02:32that it is quite similar to the example with humans. So a skill is,

00:02:37for example, if I can bake good bread. That is a skill where I say, if someone now

00:02:44prompts me to make breakfast and I have this skill of baking good bread, then that is already

00:02:50not a bad quality to produce a good breakfast, and in that case if

00:02:56someone just says make a good breakfast, I will start because of this skill that is available to me

00:03:01to also bake bread beforehand, I would set my alarm earlier in the evening

00:03:06so that I can get up at night and bake the bread, that's how I want to

00:03:11make a little bit of space, but that's how it is. So it is a bit like what you can also

00:03:14compare with human qualities first, so the human additional

00:03:18Skills that we learned in social situations, learned in the family,

00:03:23learned in friendships, through hobbies or in school, in training

00:03:27and at work, those skills enable us to perform certain specialized tasks

00:03:34effectively.

00:03:35And if someone calls you and says, 'I would like some bread,' you'll be able to

00:03:40look it up. And I want to briefly shift to technology now,

00:03:44yes, because if I ask Cloud Code or ChatGPT to bake me some bread, right now

00:03:51not much comes out. Unless I have the skill that talks to my bread machine,

00:03:56but briefly, what are such skills, yes. So skills are either

00:04:02with a point skill or they contain things like Skill MD. MD stands for Markdown,

00:04:08which is a kind of text format, where you used to say pig's card back then,

00:04:15today people say Heschzeichen, for example outline level, how to define headings.

00:04:20And then you basically write in what is the trading instruction?

00:04:25You also give an overall header, they call it Jammel-Hetter, but let's set that aside,

00:04:30there is basically a title and a description, the title could be.

00:04:35What shall we take that is nice? Write me a totally great text for social media to promote podcasts.

00:04:43So you have a title, podcast advertising. You have a description, which states, I am the great image that is capable of,

00:04:51finding exactly the right tone for the mega-successful ZinkGIF and ZinkAI podcast for promotion on social media.

00:05:00The Markdown part would then describe how such a message is structured? Yes, you had a title, you had content.

00:05:05So no title today means it is Huck, so that people stick around when they see it on social media.

00:05:11Perhaps there are also specific commands in there, like how to generate images, so with Claude I am also able,

00:05:20to provide resources in such a skill, provide AP calls, MCP calls.

00:05:27I can even provide Python code, which is then really processed rigorously and then you could take it that far,

00:05:36that if I have written this skill correctly, the system does not process it rigidly like a program code,

00:05:43but nonetheless, instead of me writing 30, 40 prompts here, give me a hook, give me a text, make me an image, upload this for me,

00:05:52Can you give different capabilities to such a skill when I interact with my AI system in the future and I write something and the system notices,

00:06:04Mark is currently writing podcast promotions on social media, he surely means this skill,

00:06:11because there is something in the description and the title that suggests he thinks this can be used.

00:06:17that it kind of gradually works through it to produce a result.

00:06:23And maybe now a small different example, before I take a breath again.

00:06:27If one thinks, okay, we all don't have podcasts, what is this example about?

00:06:31I want to take a much more striking example.

00:06:34Perhaps we all know, who work for a company, the topic of PowerPoint templates.

00:06:41PowerPoint templates are characterized by sometimes being more or less extensive.

00:06:46But they have a corporate document, they have a title, there are various pages that show, then there were transfers, graphics, charts, bullet points on title slides, subtitle slides, closing slides.

00:07:00People such stuff there are color values there might also be somewhere such pictogram databases

00:07:08where you get a whole bunch of pictograms like how do I show a tree and how do I show a dog and

00:07:14I don't know what in corporate style and it could, for example, go to Claude and say you

00:07:20run that through, look at that, build a skill, discuss with me until I

00:07:28I believe you've represented the skill well. There are test runs involved. There are

00:07:33then feedback included. I still remember, because the first time I said,

00:07:39why did you actually learn that, it doesn't look like it's supposed to. He then

00:07:44apologized. Yes, and at the end of the day, when you are finished, you have a skill,

00:07:49that is capable of making slides. They might not be at the level of the

00:07:55star designer, but I don't know if you recently saw anything from me, they don't look like they come from a star designer either.

00:08:00No comment.

00:08:01But you have it immediately in the right font, in the right layout, with the right imagery, and that is really cool.

00:08:11Especially when you also make it available to others.

00:08:15Exactly, that's another interesting point we'll go into shortly, this making available to others.

00:08:21I want to bring up another example, because you've already mentioned high sophisticated again,

00:08:25and you've even created your skill with AI again.

00:08:28Well, but if we simplify it again, it is, as I said, the skill is,

00:08:32basically, it's important to understand that it is not a tool, it is not a plug in that sense,

00:08:36and it is not an API key that I just use, but it is a bit

00:08:40about describing behaviors, what should happen.

00:08:45For example, if I say you are now the skill for a senior code reviewer.

00:08:52So a person who should look at the code again from a senior perspective,

00:08:56he has to go through several steps where he would independently read the code,

00:09:01define it in the skill. He should follow certain standards, programming standards,

00:09:05that he looks at. Then he should see that he maybe creates a test plan,

00:09:09then he should give a structured review, and then he should maybe create a risk level for any errors,

00:09:15he found during this review. That would be a very simplified depiction of the skill I could have.

00:09:23This is interesting and what I've already hinted at. This is that this skill now, because it exists in such an MD,

00:09:29or in such a Markdown file and is actually just a description of

00:09:34governance, of a behavior, is of course totally portable. That means, once I have created it,

00:09:41I can also make it available to any other model, any other bot. That I say,

00:09:47this skill is so well described and regardless of which model is working with it, it fulfills

00:09:53its purpose completely. I think that's another really interesting point because it's so,

00:09:57that's why there are already some skill libraries like this, but I also believe

00:10:02there's a whole new world beside tools, plugins, blah blah blah, all that is out there

00:10:08that there are now skill libraries where I can download good, verified, hopefully

00:10:15skills so that my AI doesn't misfire,

00:10:20at that moment.

00:10:21And today, we will again open the big security topic, as we've discussed in some previous episodes,

00:10:24but of course, also take caution with the skill download mode,

00:10:30that you are using.

00:10:31So always make sure that your skills are also reviewed again when you

00:10:33perhaps download them.

00:10:34What does that essentially say to your AI, how it should behave?

00:10:37That's what I meant before with the markdown file.

00:10:41Yes, you have a text file that describes what the thing does,

00:10:46and it's worth taking a look at it because it looks nicer in markdown.

00:10:51There are tools on the market where you can throw it in and get it nicely formatted.

00:10:55You can print it out, you can view it on the screen, it's no big deal.

00:10:59And now maybe unfortunately, something technical again,

00:11:03if you get a point-skill part and you think, 'Now I was told,

00:11:10markdown, I'll go into the editor,' you won't see anything because this

00:11:16point-skill part is actually a zip file, which means you have to rename point-skill

00:11:21to point zip, then you can unpack it, inside you will find the

00:11:24markdown files. Why is that? Well, you are delivering a container, so

00:11:30this point skill file, where various things are included.

00:11:33There are resources included, such as you can provide images, there are

00:11:38Markdown files included, these descriptions of the instructions, there is the

00:11:42header included, and so you don't have to provide everything in individual files to the system

00:11:47you can also develop it in the software, there are things like

00:11:51cursor and vintner, whatever they are called, you can drop all the

00:11:55skills in there, the point skill files, and so that you only have to drop this

00:11:59file, as I said, once and not all the little contents individually,

00:12:03it is saved by the system as a pump skill, which is actually just a zip archive

00:12:09with a fancy name for those who were paying attention earlier. What did you call the hashtag

00:12:15called? Pig Gutter? I had to quickly chime in here, I have five or six

00:12:19minutes. I learned on a typewriter. A nice lady came

00:12:24with a towel and laid it over her hands and it was still called Pig Gutter back then.

00:12:30Okay. So in our case, it was called Pig Gutter. I have a technical training. Maybe that’s

00:12:37still important for the context. Okay. Now not everyone among our listeners is always

00:12:45the one who deals with Sipfalt or other topics. Where can he find skills today?

00:12:50are there in the normal chatbots from OpenAI, from CloudCord. Are there already skills,

00:12:59that can be used, that I can turn on? To be honest, I stumbled into the topic of skills

00:13:07by trying it out. I uploaded my first skill with CloudCord without zip folding.

00:13:13I just told it, look, I need a skill, I've heard,

00:13:17they are cool and then basically asked it to help me with it and then you immediately notice they are already grabbing

00:13:25skills because you see in the chat how it states or the user wants to create a skill I

00:13:29will load the skill creator skill, and then you can have a kind of conversation with it.

00:13:35playing a long question-answer game until at the end of the day he is of the opinion that he can give me a skill

00:13:41and I promise then that it will stop and then hopefully, it will start again at some point

00:13:46I hope at least, then you'll get, if it’s a very simple skill that you

00:13:51don’t need any resources or anything for, you’ll also get the skill point MD file,

00:13:56so just this Markdown file and then it offers you, should I include this in my library

00:14:02would you like to download it, and when you download it, you’ll get

00:14:07this Markdown file and you can throw it everywhere, you can throw it in the Geminer

00:14:11Just like how you would also throw in a prompt. So copy and paste,

00:14:19name it whatever you want. However, it would then also only stay in context. So for the

00:14:24normal standard user now at OpenAI, if I would do that then it would

00:14:27essentially behave in that thread that I have open, in that case, the AI would

00:14:33actually behave like that. I think I should also mention, it also depends

00:14:35on whether you are only a non-paying user in quotation marks. Also, the

00:14:39things that are at Claude are also the pro and enterprise and such variants and

00:14:45max variants that you can buy, where you can work with the skills. But of course

00:14:50I can, and this might be important to mention as well, I can of course take skills

00:14:55in this structured form, get them from elsewhere and then also throw them into a thread like this,

00:15:01so the AI can handle better, because that is a structured

00:15:04way of how it behaved. Just important, if you are not

00:15:08going there, in case we paid at the moment, then it would only be provided to you in that

00:15:11fractal, again depending on which provider you are currently at.

00:15:15I think the skills system at OpenAI is even stronger, so they are still working on it,

00:15:19coming. That is actually the output of the cloud, I believe that is among the

00:15:22farthest with the skill topics. They introduced it relatively early on.

00:15:27They also started it with CPTGPT; if you

00:15:32create the event CPT there, for example, it also says,

00:15:36What should the agent's behavior be and so on, you can also copy these

00:15:41Markdown contents very nicely so that you basically

00:15:48have it like that, why approximately?

00:15:52You can only get the Markdown part in there.

00:15:56You can't just throw in everything, like saying, okay, I'm just going to

00:15:59throw in all the resources, you can get everything in the chat,

00:16:02but it doesn't behave quite the same way as with the skill.

00:16:05With the skill, it knows that it belongs together, it works together, it is addressed

00:16:12in the context of the chat, when I need it, while if you just copy it into the prompt

00:16:17then you just have the system instruction, so the Markdown part thrown in,

00:16:24then maybe you later throw in some files, the qualitative

00:16:27result at the end is not quite as good, but if you stick to these skills,

00:16:33say, this is how you research. This is how you prepare something for me. That is important for me to

00:16:38generate a hook. This is it, that, the other. Then the systems do not

00:16:44make a distinction, and then you can create skills, even with tools of your choice.

00:16:49You don't need to conduct the interview in Claude for a skill. You can

00:16:53simply say, dear Stretchability, dear Gemini, dear

00:16:56Mistral. I am providing you with some information here,

00:17:01structure it for me so that it is in nice Markdown; then you basically have a skill, also a

00:17:08One would say this is a work instruction, and it's somewhat similar for a career.

00:17:13Yes, I believe, to offer nothing to make it good again, to come with Gemini, I think the skills are

00:17:20closest. Yes, I think there are gems or something like that at. These are kind of role stories at Gemini, in principle,

00:17:25that are somewhat similar, where there is a role description. Because we knew that from this

00:17:29early time of prompt engineering, where it was always said, if I want to write a good prompt

00:17:35then I have to formulate my goal clearly and ideally, I also have to tell the AI buddy

00:17:41what kind of behaviors it should exhibit. And that is what is now

00:17:46actually in skills or jams or whatever you want to call it. And these behaviors,

00:17:49that has opened up. Because it's smart to out-source that. If

00:17:53it's essentially from the, from the interaction, so from the communication with the bot, which is then, in

00:18:00principle, the prompting in that moment. If I don't have to say again every time, behave like

00:18:05my uncle hasn't seen or something like that. Or am I behaving like Steve Stocks, or am I behaving

00:18:11like this or that, or am I behaving like a good programmer, or am I behaving like a teacher

00:18:14for languages or something like that. So I don't need those kinds of things to prompt over and over again,

00:18:20Essentially, we're talking about the idea today that it means,

00:18:26I have to prompt less to get better results because the AI understands through the skills,

00:18:35that I either provide in the moment or as a file somewhere,

00:18:40where she can then still access it or smartly access it herself,

00:18:44Because of course, when you offer thousands of skills or 100,000,

00:18:50Providing skills is, of course, also smart, so that the career doesn't load all skills

00:18:53but just an image, as you described, simply loads the skills as well,

00:18:56that are necessary to inherit a certain task well.

00:19:01One must not forget also the selection of a skill, but rather the invitation

00:19:06into the system.

00:19:07The system then notices, okay, there is a skill that I need, or is there even

00:19:11one.

00:19:12I load this in, that’s already the so-called tokens for your request. This means,

00:19:18it makes sense, if you want to operate more professionally,

00:19:23maybe also to start with something like orchestrator skills, that you say, okay, I have

00:19:28a skill that basically selects the appropriate skill and has its own set of rules so that you

00:19:33don’t just have a list of 1000 skills, but the orchestrator essentially says,

00:19:39okay, fine, we are developing, maybe it will lean more towards this guild,

00:19:43a text will be created and maybe this will be the guild. But those are, I’d say

00:19:47advanced technologies. And as you mentioned earlier with discussions,

00:19:52like those jobs and so on, we talked about it in my episode, I had built this advisory board

00:19:57back then with N8N. I then went and threw this advisory board into the

00:20:02AI and said, can you please provide this as a guild? And

00:20:06it did exactly that. It created an orchestrator skill, which is the one that defines

00:20:13the topic, selects the experts, and then interviews the experts accordingly.

00:20:19All the information was processed in there. It ultimately became a huge

00:20:23part of skills at the end of the day, but only partially, due to zip, compression,

00:20:29but that’s another topic. And then you can give it this, and when you tell it,

00:20:34‘Hey, I need the advice of the experts,’ then the team goes and says, ‘Advice of the experts,’

00:20:40experts.

00:20:41I’ll look for it and check it out and then you see it gets really lively in the chat because then

00:20:47he starts and says, yes, now Elon has to be with Steve, with the Osnack, with whom

00:20:51always discussing that, and in the end, something comes out.

00:20:55You're thinking, okay, we talked about this topic last time when we discussed it back then,

00:21:00does it actually make sense to give the different people with

00:21:03or not, but I still think that it's a very cool result

00:21:08that is much more than if you don't have all this contextual knowledge of the person from them

00:21:13you had given. And maybe, with all the nerdy talk today, which might

00:21:20be a bit difficult for one or another, I just realized how

00:21:24often I've completely lost the context in my head with Skill and Skill MD

00:21:29could. An important story that each of us can tell is to take a piece of paper and

00:21:34a pen and according to the old process knowledge, write down what you are doing, what you have written down,

00:21:40because when you document that for some time, you're getting into your topic, what occupies

00:21:45you constantly, what are the topics that may just annoy you, yes, because I have to

00:21:49here again, no idea, end of the month, collect all the receipts from the bank and

00:21:55I sort them by the term ABC and for taxes, it's important with the VAT

00:22:00as well. You could also consider, I say, you know what, I'll make myself

00:22:05a skill. The thing should take the PDFs or whatever I get from the bank in statement extracts,

00:22:10read them out, create folders, place the stuff there, and then you already have

00:22:17written a work instruction that the agent processes. It helps you. Yes,

00:22:23it frees you from the terrible things and the annoying things, and you can assume that if ever

00:22:29a mistake should happen, you adjust the skill and from then on it will work better in the future.

00:22:35As bad as today, AI will never be again. And the point you had in the preparation,

00:22:41when you throw over the fence, I think it's worth mentioning at this point again,

00:22:46that these skills not only serve to give you a better result.

00:22:52So I distribute the bank bookings into a folder for accounting or I create

00:22:59a PowerPoint, but also to guide him on how he goes about this, so that he keeps following this path,

00:23:07step 1, step 2, step 3,

00:23:13and that he also documents what have I just done? Because imagine, you give him the

00:23:21task to reorganize my hard drive.

00:23:24I also mentioned this recently with Claude Co-Work, you have a 4TB drive

00:23:29and he is supposed to come up with a folder structure and naming conventions, and if he does this

00:23:35without you telling him, then he will do it, but eventually you will notice,

00:23:39why are the folder names different now and the file names have changed, why?

00:23:44He has changed the scheme, and that is simply because

00:23:48at some point the context was so full that he no longer knew, like a goldfish, what was

00:23:53happening just two minutes ago? But if you tell him, create a memory, save your

00:23:58decisions, save special cases, note the follow-up questions, then at least you have more

00:24:04security for this work case, 100 percent doesn't exist, but more security that he stays

00:24:11on track. And if you store this more generally, then you quickly return to

00:24:16our favorite crab OpenClaw and Co. That remembers not only step by step,

00:24:22but in a more general way. But I know. Yes, yes, all good. But then, because you just brought up a

00:24:28good point. Now you've mentioned memories. So now for example, I think,

00:24:32we can make a little distinction here. Just a Skill and Memory Files. No,

00:24:37because of course with OpenAI, I don't even know if that is included in this $20 paid version

00:24:42and not in the lower ones. Essentially, there you can also say Open-R, remember

00:24:49I want a text, for example, to be written with emojis. When

00:24:54I say, write me the next link in the article, then use some emojis because

00:24:59that’s my style. What we are discussing, that I discussed with my Chatchapity variant

00:25:04is partly stored in this memory file, because it knows,

00:25:09for example, that I am discussing such an episode, that I will soon do an episode about some topic

00:25:14about the market, then ChatGPT knows the context from this memory file. Why,

00:25:21what I mean with my podcast and what I then do with it. That’s just a bit

00:25:25different than skill. Skill is more clearly defined behavior in my opinion

00:25:31and should still be established, so to set the governance. Memory goes a bit beyond that.

00:25:34I think that’s a slight difference that isn’t always so clear,

00:25:40which should be distinguished.

00:25:42Between a memory case that knows all context and more about my personalities.

00:25:48And skills that actually describe potential specialization abilities at the moment.

00:25:57So I would roughly distinguish that, because I think it's also important to understand.

00:26:01Memory overall file might be even more powerful, but it will also become

00:26:08larger and could also have more errors compared to perhaps a very specialized skill.

00:26:13the file that I can also port when I can load the motor and other things like that I would

00:26:16then distinguish clearly and I have that it is also not like you have to remember that

00:26:23the thing if it now structures my hard drive has to remember for eternity

00:26:27what the file names are, while whether I eventually, I don't know, drink my beer more on Sundays

00:26:33and accordingly more recommendations for my future brewing, could possibly

00:26:39depending on the situation be more important to me than, as I said, the mentioned file folders from the backup hard drive.

00:26:45What I believe should stay with you all from this episode today is,

00:26:51that if you've never dealt with prompt engineering and so on, basically,

00:26:56just write it down and it will produce results, results are good. Then that's fine. Personally, I have now

00:27:02also come to the realization that I can find a more sustainable solution with these skills,

00:27:09because it definitely provides more fun, joy, and better results to engage with the topic of

00:27:17skills. Yes, I am more of a skills person, yes, others might be more the

00:27:23jams person, and others I don’t know, whatever the systems are called, but because

00:27:28it's open source, and I will say, in various software products it is also

00:27:37supported, maybe today in software movement, but who knows, maybe some

00:27:42requirements engineers, tax consultants, I don’t know what else, people who

00:27:47work in their systems with one system, and that system maybe internally

00:27:53uses different skills to provide features or acts as a partner

00:27:59or whatever is available, and I would like to promote that.

00:28:04What I mentioned earlier with paper and pen, also write down what you do,

00:28:09and what you have written down. First of all, to find out for yourself what the

00:28:11annoying tasks are, because we should look at those first and then also

00:28:15think about how to transfer it into markdown, then it already looks nice, yeah,

00:28:20then you can also give it to your colleague as a vacation substitute or at least

00:28:25see how the machine handles it, and yes, I think that would be my message that I

00:28:31would really like to share with everyone. I think that's good, I would go one level of action higher

00:28:36and say, actually it's also a bit, if you want to make a very

00:28:40light architectural image, to say how these different things

00:28:45are, I would say then this skills layer is clearly one, this behavior and reward layer, where I apply different behaviors,

00:28:54secondly, different rewards according to the norms, when you look for errors or similar things,

00:29:00I can place that in the skills.

00:29:02Then you have the next layer, something like, we often talked about the MCP or the Tool Layer.

00:29:09If you will, the AI then accesses its skills, its behavior, how it should behave, to possible tools, whether that is web browsing, whether it's 3D rendering, or whatever, then executing those through a tool.

00:29:27Baking.

00:29:28Bread baking.

00:29:29The bread machine in case of emergency, that can then be, the skill would be that the bread

00:29:35must be kneaded with love again, and other things about what you do.

00:29:40And then I essentially have the third layer for me, the One Time layer, where that

00:29:43actually happens.

00:29:44There we are then at the layer of the actual, whether that's a AI model in a

00:29:51context window only, whether that's an open cloud in the one time, on which these things

00:29:56can run in that moment. I think that's a little bit, if you want to place it electronically.

00:30:00This skill against an AI model, or an agent network,

00:30:06or a bot like an open cloud that tries to independently continue to do things, so that you can somewhat differentiate.

00:30:11Skills, behaviors,

00:30:16to the tooling layer and then you have the one-time layer where something happens, where also things

00:30:21are executed, which then essentially uses the available tools as well as skills.

00:30:27And all of this is supplemented by the actual prompt that comes from us, which contains the

00:30:32task I want to have done. I find it really difficult today, I think,

00:30:39to follow us. I'm never sure. I'm very curious about the comments. But also something like that

00:30:45needs to be addressed because we had, if we briefly touch on the last

00:30:49Yes, listen to the episodes. Greetings again to Alex, Klaus from last time, where we said,

00:30:55it's not just the technology that there is a new model. It's not just the human who has to figure out

00:31:02how to use this new world. Yes, we are now writing the skills, for example. It’s

00:31:06also the complete interaction, when you think about how things work together, because now

00:31:12I write a skill that perhaps deals with a topic that I work on, but if I

00:31:16include two or three things from my colleague, I am practically cutting work packages. From

00:31:22that perspective, it is a very powerful world, and that's why I found it very important that we talk about

00:31:28this technological basis, skills today. Maybe a funny anecdote,

00:31:34a colleague of mine today was told by Claude that Claude couldn't solve a problem.

00:31:40That’s good, that shows that the human is still in the loop. I found that very funny,

00:31:45yes, so from that side a little humorous activity. You’ve already hinted a bit,

00:31:51pre-discussion, about what we will talk about next time, and I would like to tease a little

00:31:55for next time. Could you maybe give us two teaser points,

00:32:01before we wrap up for today? So if we take a look,

00:32:06if we quickly glance over the last few years of AI,

00:32:10we have come from the Ahamut, to JGPT, we’ve gone through prompt engineering,

00:32:15we’ve gone through possible workflows, through workflow engines, to now

00:32:21in 2026 nearly self-working AI bots that are out there,

00:32:29running around the internet in crab forms.

00:32:32And we are slowly entering a world where it’s not just one of these bots,

00:32:37where I don't just have one bot equipped with skills and tools and you name it,

00:32:41but certainly in my orchestration layer, I then have many of such bots

00:32:50out there.

00:32:51And what Mark and I have observed in the past weeks is a trend

00:32:55on the internet, that some of these colleagues, who also run 15, 16 AI agents in parallel

00:33:02are starting to equip them with a so-called gaming UI, where they can then observe

00:33:09and also orchestrate what these AI agents are doing, because you might not want

00:33:13to give everything via prompts anymore, but perhaps want to display a kind of cafeteria

00:33:20where the bots that are currently busy meet during their idle time and

00:33:26I can then decide as a human again in Belube that I say,

00:33:29You know what? If the watcher is hanging out in the cafeteria all the time and using idle time,

00:33:34then I’ll kill him now, because I’d rather wait for a new task to come.

00:33:39No, of course, I’m not going to kill him, I really won’t kill him.

00:33:42Not kill, man, you should assign the tasks, so.

00:33:44Yes, yes, yes, exactly.

00:33:45Ah yes, but as I said, there are definitely things, now, that you can do outdoors

00:33:50really to conduct such an Orkestation, building gaming UIs and we will or gaming

00:33:57similar UIs and we would really like to talk about that in a complete episode and I'm very much looking forward to it because it's one of my favorite topics.

00:34:04And Mark is also very eager for this episode and we've been planning it for ages, I think, it's been ages in this AI.

00:34:11You've nearly, well, for about three months, and we've been planning it for ages and now we're doing it.

00:34:18And if you don't want to miss it, tune in next time. We've never hinted so strongly beforehand that it will happen.

00:34:24But this time we are doing the fire-sage, Jens.

00:34:27Yes, next time we will do the episode.

00:34:29Let's see if it works out, if you liked the episode,

00:34:31then leave a comment, stars,

00:34:34whatever is possible on your platform.

00:34:37And we look forward to the next time at Age of Agents.

00:34:42Until then.

00:34:43Ciao.

00:34:44Until then.

00:34:46Welcome to Think Different, Think AI.

00:34:49The podcast by Mark and Jens.

00:34:52Two tech-loving minds who not only talk about artificial intelligence but live it.

00:34:58Here you'll find clear categorization, real practical insights, and a fresh perspective on what's possible.

00:35:05Understandable, critical, and always with a wink.

00:35:08HDI to ponder, to chuckle about, and above all, to discuss.