The AI Skill That Actually Matters Isn’t Knowing Tools — It’s Building Workflows

Learning how to build an AI workflow is worth more than mastering any individual tool, and the reason is simple: tools change every few months, while the ability to chain them together doesn't. Atlantis University's AI workshop made this point structurally rather than rhetorically — three separate sessions, three different presenters, and each one built directly on the last. By the end, the room had a complete customer acquisition system.

What was actually demonstrated

Read the sessions in sequence and the design becomes obvious:

Session 1 — Content. Extract a visual identity into structured data, install it as a persistent design system, and generate on-brand social content from any article on demand.

Session 2 — Destination. Take a source — a video, a book, a set of documents — extract its knowledge, generate a build specification, and produce a functional website with working tools.

Session 3 — Conversation. Add an AI chatbot to that site, connected to a model and grounded in a knowledge base, so visitors get answers around the clock.

Professor Rodolfo Capdevilla framed the connection explicitly when he opened the third session:

"You create your content, and you attract people to your business. But then they need to land somewhere. That's what Julian taught us — the landing page. Now these landing pages need a few tools. We're going to add one of those."

Three sessions. One system: attract, land, engage.

Why single-tool expertise ages badly

The clearest articulation came from Julian Borge:

"The real skill with artificial intelligence is not knowing the different tools. It's creating the whole workflow. And it applies to any single aspect of a business you're creating. If you're an expert in lead generation, you have to create a whole workflow process. If you are an expert on sales operations, you need to create your whole workflow."

Consider what this means practically. The tools named across the workshop — Claude Projects, NotebookLM, ChatGPT, Lovable, Google AI Studio, Base44 — will look different in eighteen months. Some will be gone. Others will have absorbed each other's features.

What survives is the pattern. Extract knowledge from a source. Convert it into a specification. Generate the artifact. Iterate on the errors. That sequence held across all three sessions, with different tools in each slot.

The four moves that recur

Pull the sessions apart and the same four operations appear repeatedly:

1. Ask the model to write your prompt

The most reusable technique from the entire event, and it appeared in every session independently.

"Do not try to create professional prompts. Just ask the LLM for the prompt."

John Castaño screenshotted an interface and asked ChatGPT to write the instructions for it. Julian Borge asked ChatGPT what questions to ask NotebookLM, then asked it again to build a website specification. Rodolfo Capdevilla generated his five chatbot prompts the same way, then split them into stages.

Nobody hand-wrote a complex prompt. They described the outcome and let the model construct the instruction.

2. Convert unstructured material into structured form

The style extraction into JSON. The video into an organized knowledge base. The business plan into chatbot grounding.

The intermediate artifact is what makes the next step reliable. Prose descriptions get reinterpreted differently every time; structured data doesn't.

3. Install it as persistent instructions

A Claude Project. A custom GPT. A chatbot knowledge base. Same principle: configure once, use indefinitely.

This is the difference between level-one AI use and level-two. One-off requests versus a system that retains context.

4. Iterate on errors rather than restarting

Capdevilla's chatbot failed on a previous attempt. His advice: screenshot the error, hand it back, ask for a fix.

"There is an iteration process, a polishing process, after you create this. Most likely you're going to have to do that."

First output is a draft. Treating it as a verdict is what makes people conclude the tools don't work.

Which tools to actually pay for

Useful guidance, and notably candid for an event with a commercial purpose:

"Before paying for any AI tool, test it and try it. Do not throw money into the garbage."

The pattern among the presenters was consistent. They paid for the models they use daily — typically one or two subscriptions in the $20-per-month range — and used free tiers for everything else. The complete website in session two was built on free NotebookLM and free Lovable. The chatbot in session three ran on a free Gemini API key.

A workable rule: subscribe to the general-purpose model you'll open every day. Use free tiers for specialized tools until one of them becomes part of your routine.

You don't learn this by watching

The most repeated instruction of the day, from Professor Alex Lima:

"You do not learn AI by talking about it. You don't learn AI by reading. You don't learn AI by watching videos. You learn artificial intelligence by doing stuff."

Every session was structured to enforce it — presenters continually asking the room to confirm they were following along, building live rather than showing finished results, and distributing step-by-step guides afterward specifically so attendees could rebuild everything themselves.

And Borge's own timeline is the strongest evidence for the approach. Ten months before presenting, he couldn't build any of it. What changed wasn't watching — it was, in his words, a lot of testing and error over an extended period.

Frequently asked questions

What is an AI workflow?

A sequence where each tool's output becomes the next tool's input, producing a finished artifact that no single tool could deliver alone.

How many tools should a workflow use?

As few as the job requires. The workshop's website workflow used three. Complexity for its own sake makes workflows fragile.

How do I start building one?

Pick a task you repeat. Map the steps you currently do manually. Identify which steps a model could handle, and connect them.

Do I need to pay for tools to build a workflow?

No. The complete website and chatbot demonstrated at the workshop were built on free tiers.

What happens when a tool in my workflow disappears?

You swap it. That's the entire argument for learning workflows rather than tools — the sequence survives the substitution.

How long does it take to get good at this?

The presenters' own answer: months of deliberate practice, not weeks of watching. But the first working result comes on day one.

Conclusion

Knowing how to build an AI workflow is what separates people who use AI from people who get compounding returns from it. Individual tools produce individual gains. Chained tools produce systems.

That's also the honest answer to the anxiety underneath most AI conversations. The concern isn't really that a tool will replace you — it's that someone who has built systems around these tools will outpace someone who opens a chat window occasionally.

The gap is closeable. It just isn't closeable by reading about it.

The Master in Artificial Intelligence at Atlantis University is built entirely on hands-on practice and adapts to your background — healthcare, business, technology, government or military.

Become a student at Atlantis University