How I move from “AI that executes” to “AI that challenges me.” The prompt and process that actually change how I build.

I’ve seen AI generate files I’d print and frame. I’ve also seen it break contexts by being too compliant. The difference isn’t magic: it’s how you talk to it. I’ll tell you the change I made: from asking it to execute to asking it to challenge me.

A real change: when the AI left me speechless

At Western Union, after defining an initial collection of tokens, the AI refined the set and generated an SCSS so clean that I went straight to the CTO: “This needs to be printed and framed.” Specifically: it eliminated 27 duplicates, grouped semantic tokens by intent, and proposed a spacing scale that I later adopted in all projects. Learning: the value wasn’t the code, it was the conversation that forced us to decide what was immutable.

I’ve spent years reviewing and teaching token files — in companies like Amazon or Western Union, with large teams, analyzing them in courses and talks — and I had never seen one so polished, just like that, in one go.

Vulnerability: what really worries me

Yes, it’s scary that now someone with less experience can set up a system in a fraction of the time. My response is not to compete against AI in execution, but to lead context, criteria, and agreements. That doesn’t self-generate: it’s designed by talking.

When AI fails

The Achilles heel of AI? If it lacks context, it can fail miserably. Early versions of GPT were terribly compliant: they always answered something, even if it didn’t make sense. That’s where our value still lies: knowing how to say “not this way,” demanding it asks us more and adds depth — not accepting the first draft or the easy answer.

Signs that the AI is being compliant (and how to stop it):

It gives you long answers without asking for more context -> Answer: ask for 3 questions before proposing solutions.
It returns perfect code but disconnected from the problem -> Ask for “rationale, trade-offs, and risks.”
It says “yes” to everything -> Ask for counterexamples and a quick stress test.

The simplest but most powerful prompt I use

I’ll give you the prompt that has given me the best results:

You are an expert in your field. From now on you will be my advisor on \{specific topic\}. Don't be compliant with me: question my assumptions, ask me questions before proposing solutions, and raise risks and trade-offs.

Let's work in maximum 10-minute iterations and finish each round with:

1) suggested decision

2) risks

3) next question

My 5-step flow

I define the context in 5 lines (who, for what, constraints).
I ask for 3 questions and stop (if it doesn’t ask, I retry).
I demand 2 alternatives with trade-offs (not a generic “best practice”).
I validate with a small example (not full specs).
I summarize what I’ve learned and use it as a prompt for the next iteration.

The key: it’s not magic or flattery — it’s opening a demanding conversation dynamic, back and forth, demanding value and refining together.

My practical advice

Dare yourself: start a basic project of your own from scratch with the help of AI, tinker, break it, iterate, and when you finish, ask the AI to summarize what you’ve learned and start again with that summary as the initial prompt.
The second, third, and fourth versions will be better and more yours.

And you?

Do you dare to let AI challenge you — not just execute tasks for you? What was the last thing the AI taught you or that you would never have done without it?

#AIUX #DesignSystems #Solopreneur #ProductThinking

Creativity isn’t replaced by AI. But AI, if you understand it, forces you to be more creative than ever.