AI Workshop in Târgu Mureș, August 11, 2026: prompts, context and knowledge

Through Dalbe, I regularly organize AI workshops for entrepreneurs from Târgu Mureș, at Metamorfozum Concept, 3 Matei Corvin Square, in the city center. On August 11, 2026, I started a new one, intentionally split into two days. I knew from the start that the second day would move towards more advanced things, skills, agents, and automations. I wanted the first day to set the basics straight, before all of that.

One of the participants, Adorjáni Noémi, partner at OnLike Agency, noted down a single idea at one point: prompt + knowledge files. That's it. But that's where what we were trying to build throughout that whole day was seen most clearly.

Because it is very easy to open ChatGPT, Gemini, or Claude, get a good answer, and have the impression that you already understand how AI works. The problems appear later. When you give it information that maybe you shouldn't give. When it answers you very confidently, but wrongly. When you use the same prompt in two models and get completely different results. Or when you have ten conversations, three GPTs, two Projects, and you no longer know where you put the information you needed.

The entire first day was, in fact, the journey to that idea Noémi noted.

Why we started with the basics

We started a bit with history as well. DeepMind has been around since 2010, long before the general public talked about generative AI on a daily basis. Meanwhile, OpenAI, Anthropic, and many other labs and companies appeared, pushing the field in different directions. However, I didn't want to turn the workshop into a lesson on the chronology of AI, but to show how much the models have changed in a relatively short period.

The first GPT generations were much more limited than what we use today. The evolution did not come from someone simply adding a "logic module" or a "math module", but from larger models, better data, better training methods, instruction tuning, human feedback, tool use, and other techniques that made the systems much more useful in practice. Beyond all the differences between products and generations, the important idea for the workshop was simpler: models are very good at language and can formulate extremely convincingly even something wrong.

This is one of the ideas I try to repeat at every workshop: a very well-formulated answer is not automatically a correct answer too.

We also compared ChatGPT, Gemini, Claude, DeepSeek, Perplexity, Manus, and Grok, but not with the idea of finding a "winner". Each has its own strengths, its own limits, and a different way of approaching certain tasks. For me, the practical conclusion is that it makes no sense to look for a single AI that does everything perfectly. It's more useful to understand which tool makes sense for the problem you have in front of you.

We also talked about Anthropic's very explicit approach around Claude's rules, safety, and behavior. Gemini has the advantage of increasingly tight integration with the Google ecosystem, Perplexity is interesting when you want to start from search and sources, and ChatGPT comes with its own ecosystem of Projects, GPTs, and tools. We didn't want to get into brand wars, but to understand that the same prompt can have very different results depending on the model, context, and available tools.

Context helps, but not just any context should be given to AI

We checked the settings in ChatGPT together, and from here we got to privacy and GDPR. If you use AI for work, sooner or later the temptation arises to copy a contract, a client list, an export, an internal document, an email, or data from a CRM.

Here the question is no longer just whether the AI can help you. More important is what information you give it to help you and whether that information really needs to be sent to an external service. We discussed the difference between public information and internal data, what data is truly necessary for the task, and whether certain information can be anonymized or removed before upload.

Context helps the AI work better. But that doesn't mean every context has to reach the AI.

Beyond all the new features, we still remain in an area where the model needs information to produce a good result. The tool can be very good. That does not mean we have to give it access to everything.

Normal chat, Project, or GPT?

Only after these basics did we specifically get to ChatGPT and the difference between a normal chat, a Project, and a custom GPT. These are three things that many users still mix up.

A chat is good for a one-off conversation or a clear task. A Project starts to make sense when you constantly work on the same topic and want to keep instructions, files, and relevant conversations in one place. A custom GPT makes sense when there is a repeatable use case and you want to define a set of instructions, a behavior, and information sources for it to use constantly.

We also checked image generation, then moved on to GPTs. I showed mine, explained how to build one from scratch, and we checked a writing GPT made by one of the participants.

And then Noémi wrote: prompt + knowledge files.

Exactly from here begins, for me, the transition from a generic chatbot to a tool adapted to a real activity.

The prompt is only part of the result

You can have a very good prompt. But if the model knows nothing about your company, your products, your customers, or the way you communicate, there is a pretty clear limit to the result it can produce.

If you tell an AI "write a Facebook post", you will get something. Maybe even something decent. But the model doesn't automatically know who the audience is, what the brand tone is, what we said previously, what expressions we avoid, or what promises we don't want to make.

The moment you give it a brand guide, communication examples, information about the audience, and some clear rules, the same model can produce something completely different. Not necessarily because the prompt has radically changed, but because the context has changed.

And here the interesting part begins. A good prompt matters, but the prompt doesn't exist alone. It also matters what information the model has available the moment it starts working.

The same idea, everywhere

The idea then repeated itself in almost every tool we went through. We showed Google Gemini, Nano Banana, and NotebookLM, where you can start from your own sources, documentation, PDFs, and internal materials, instead of letting the model rely solely on the information it already has.

Here we also touched on RAG, without turning the workshop into a technical course. It made no sense to get into embeddings, vector databases, or architecture. The concept mattered more than the implementation: instead of the AI relying only on what it previously learned, we offer it the relevant information exactly when it needs it.

Practically, it is the same idea that Noémi had noted, just in a more technical form: prompt + knowledge files.

We also tested deep research and the difference between asking what the AI knows about a subject and letting it search, compare sources, and build a documented answer. For business, the difference can be big. It's one thing to ask for ideas. It's another to analyze a market, a competitor, prices, legislation, or trends, where information changes over time.

And here I insisted on one thing: the fact that the AI can do research does not mean sources no longer need to be checked. Automation reduces some of the work, but does not eliminate critical thinking.

From presentations to a rock song

We also went through Google Vids, HeyGen, and Gamma for presentations and visual content, each with concrete examples and how they can be used in a company.

The most fun moment was with Suno. We wrote lyrics with ChatGPT, then took them into Suno and a whole song came out. Their 5.5 engine is surprisingly good at epic rock.

But even here the process was the same: what I ask of it, what information I give it, what style I want, what examples I provide, and what I modify after the first result. AI does not eliminate the creative process. It changes it.

Why a brand guide can be worth more than 20 more prompts

The same idea came back to content creation as well. If you periodically use AI for the same company, it starts to become absurd to explain the same things in every conversation: who we are, who we address, how we communicate, what expressions we avoid, what an article looks like, how we formulate a CTA, and what services we have.

All of these are reusable context.

That's why we talked about using a brand guide as context and knowledge. When the AI has access to stable rules and relevant information, we no longer have to rebuild the same foundation every time.

Here we were already preparing, without going very deep, one of the main ideas of the second day: if you repeat the same rule ten times, you probably shouldn't manually write it every time anymore. You should start transforming it into an instruction, a document, a skill, or another reusable component.

The first day was not about tools

Towards the end we went through Projects again and started talking more about Claude, about its features, and only lightly touched on skills and Cowork, without going deep into them yet. Intentionally, because that's where the second day was going to start.

At the end of the first day, we already had a pretty long list of tools and concepts. But my conclusion was not that we need more tools. We already have too many.

The problem is knowing what information the AI needs to transition from a generic answer to one you can actually use. A good prompt matters, but the prompt doesn't exist alone. Instructions, examples, documents, sources, project history, brand rules, and the limits we set for it matter.

This is exactly what Noémi captured in those three words in her notes: prompt + knowledge files.

And from here the second day was about to start. Because the next step was no longer to learn yet another tool. It was to see how we transform context, rules, and repetitive tasks into something reusable and how we start linking all the pieces together through Projects, skills, Cowork, agents, and automations.