Day 2 of the AI workshop: two reactions, one small group
On August 12, we held the second day of our AI workshop, this time at the OnLike Agency office.
The workshop was attended by Pál Alpár from Gheorgheni, invited by Adorjáni Noémi from OnLike Agency, our collaboration partner.
I intentionally wanted a small group. In such a format, I can better see where questions arise, what sparks interest, and where I need to change the pace. I don't have to mechanically follow a set of slides, but I can adapt the workshop to what is happening in the room.
The second day was designed as the more advanced part.
If on the first day we built the foundation, now we moved towards prompting frameworks, skills, Cowork, agents, MCP, automations, local AI, and some practical examples where we tried to connect all these concepts to real work.
We started simply: how to reduce the things the AI has to guess
We started with prompting.
We went through some known frameworks, such as CO-STAR, SMART, role-based prompting, one-shot and few-shot prompting.
But I didn't want the participants to go home with yet another list of acronyms to memorize.
For me, all of these are essentially trying to solve the same problem:
how do we provide the AI with enough context so that it has as few things to guess as possible.
We used a very simple example.
Instead of just asking:
Give me the quote of the day.
I also gave it the format in which I want the result:
Quote of the day: "Simplicity is the ultimate sophistication." — Leonardo da Vinci
Just one example.
The model immediately understands the desired structure.
This is one-shot prompting in its simplest form.
From here we discussed the situations where one example is enough and those where multiple examples, meaning few-shot prompting, help the AI better understand the pattern.
We also got to a much more mundane, but very practical detail: copy-paste.
When moving texts between ChatGPT, Claude, Word, a CMS, or a newsletter editor, we can take unnecessary formatting and HTML with us.
The text might look fine on the surface, but junk remains in the code.
It is a small detail, but exactly the kind of thing that starts to matter when the AI is no longer just an experiment, but enters a real workflow.
From a prompt to something we can reuse
From here we moved on to skills.
Together we built a marketing audit skill.
This is where I think one of the important shifts in how we use AI occurs.
If we explain the same methodology ten times, it probably no longer makes sense to write the same prompt ten times.
We can turn that process into a reusable skill.
We define what needs to be analyzed.
We set the criteria.
We define the structure of the result.
We test it.
We improve it.
And next time we don't start from zero.
The same principle can be used for marketing, SEO, sales, or project management audits, or for company-specific internal processes.
At that point we are no longer just talking about "a good prompt".
We start building a way of working.
Then came Cowork and the pace changed
A significant part of the day was dedicated to Claude Cowork.
We discussed how it can work with skills, files, connectors, MCP, and multi-step tasks.
For me, the interesting part about Cowork is not that we have yet another interface to talk to an AI.
It's the fact that we are starting to transition from:
"I ask you a question and you answer."
to:
"Here is your objective, context, files, and tools. Work on the problem."
And here the workshop became much more interesting than any demo I had prepared.
Noémi started testing the skills and Cowork on a real case and, right during the workshop, created a document with lead generation opportunities for Dalbe.
I liked this more than any fictional example I could have prepared.
Because in that moment she was no longer testing a tool.
She had already started thinking:
"Where can I use this in my work?"
Pál Alpár had exactly the opposite reaction.
He connected his email and calendar and immediately started testing everything he could.
Email drafts.
Information analysis.
Calendar-based plans.
Questions about the data Cowork had access to.
Every answer immediately generated another question.
"Can it do this too?"
"What if I connect it here?"
"Can it analyze this?"
"Can it make the plan based on the calendar?"
At one point he was, honestly, like a kid in a candy store.
And I thought it was a very good moment because it showed two completely different ways of discovering the same technology.
Where does automation end and AI begin?
From Cowork we naturally arrived at automation.
We discussed Make and n8n and the difference between a predictable workflow and a system that needs to interpret information.
In n8n, for example, we can define:
form completed → create contact → transform data → send notification → save result
There is a trigger.
There is logic.
There are steps.
The process can run by itself in the same way every time.
We don't need an AI to "think" between every two nodes.
AI becomes useful where interpretation is required.
For example:
new lead → AI analyzes the company → classifies the opportunity → writes a summary → the workflow saves it in the CRM
For me, the question is not:
"AI or automation?"
But rather:
"Where do we need judgment and where do we just want things to run predictably?"
In many processes, the right answer will be a combination of the two.
Here we also discussed Hermes, agents, connectors, and MCP, not to go very deep into the architecture, but to understand the direction.
The prompt tells what we want.
The skill can define how we work.
The connector provides access to another system.
MCP can standardize how the AI uses different tools.
And an agent can combine several of these to achieve a result.
We also built a website, but that doesn't mean the website is done
We also did a website creation exercise.
We worked in parallel with Claude and Lovable via MCP, building a landing page.
I wasn't necessarily interested in which tool generated something spectacular first.
It was more interesting to compare how they interpret the same requirement and what happens after the first version.
Because today we can generate a website very quickly.
That doesn't mean the website is good.
We still have to ask ourselves:
What is the objective?
Who are we building it for?
Is the message clear?
Does the structure make sense?
Is it responsive?
Is it accessible?
Is it SEO correct?
Can it be maintained?
What happens after the demo?
AI can enormously accelerate execution.
But if the problem is poorly defined, we just get to the wrong solution faster.
AI doesn't strictly have to run in the cloud
I also wanted to show a slightly more technical area: local AI.
We installed and ran Gemma through Ollama, including the variant:
ollama run gemma3:12b
We went through the installation steps for both macOS and Windows.
Not because everyone should give up ChatGPT or Claude.
But because I wanted it to be clear that an AI model can also run locally, on your own computer.
For certain projects, this can become relevant for reasons of privacy, offline work, experimentation, or greater control over the environment.
Of course, the fact that the model runs locally does not automatically mean the entire system is secure.
But it's important to know the option exists.
A restaurant showed us very well what context means
At one point, a question came up about Leadomator and how much AI can already help in researching companies and potential clients.
Instead of trying a sales case directly, we chose something much simpler.
Restaurants.
The first request was roughly:
Check all the restaurants near 12 Retezatului Street in Târgu Mureș, within a maximum radius of 1 km, and recommend a restaurant.
The result was OK.
But generic.
Then we changed the request.
I said that I'm a person who doesn't like fast food, that I want the menus analyzed, and that every restaurant needs to be evaluated from this point of view. I asked for the results in a table and the Google or Tripadvisor ratings where they exist.
Suddenly, the answer became something you could actually use to make a decision.
The model was the same.
Nothing technical had changed.
The question and the context we had provided had changed.
And exactly the same principle applies in sales.
"Find me potential clients" is almost useless.
It is much more useful to define:
what kind of company we are looking for;
in what area;
what size;
what technology they use;
what signals indicate a need;
what we want to exclude;
and by what criteria we evaluate the opportunity.
In the end, two people went in two different directions
At the end of the day, Noémi said something very honest:
It was a lot of information and she needs a little time for it to settle.
I agreed.
When you transition from ChatGPT used as a simple chatbot to skills, Cowork, agents, MCP, automation, and local models, dozens of directions in which you could go appear very quickly.
We agreed she would continue testing the built skills and, if questions arise, we'll discuss.
But you could already see she had started thinking about opportunities.
About what she could build.
Where she could use skills.
What processes she could change.
Pál Alpár, on the other hand, didn't seem to need that processing time.
He was already there with his email connected, calendar connected, and was still testing.
One wanted to let the information settle.
The other wanted to open every box on the shelf.
And I think both reactions are correct.
Some people need to understand the system first and only then introduce it into their work.
Others learn by connecting things and immediately seeing what they can do with them.
For me, that was the best part of the day.
Not the fact that we went through CO-STAR, skills, Cowork, MCP, n8n, Gemma, or Lovable.
But the moment the participants started to see for themselves where they could use all of this.
That's what I try to achieve through our workshops.
Not to go home with yet another list of AI tools.
But for you to look at a process from your own activity a few days later and have the thought come to mind:
"Wait a minute. I think I could do this part differently now."