Most people try to make AI give better answers.
For a long time, I did the same.
I improved my prompts, experimented with different models, and spent hours refining my research process. The responses kept getting better, but something still felt wrong.
Eventually I realized I wasn't optimizing for truth.
I was optimizing for answers.
This article is about the moment that changed how I think about AI—and why I no longer want an assistant that simply agrees with me.
I want one that challenges me.
For years, my research process looked like this.
I would watch YouTube videos, browse TikTok, search Google, compare articles, read Reddit threads, and eventually ask ChatGPT to summarize everything.
Depending on the decision, I could spend one or two days researching before feeling comfortable enough to move forward.
The problem wasn't the time.
It was trust.
The internet has become incredibly good at looking convincing.
The first Google result isn't necessarily the most accurate one. Sometimes it's simply the page with the best SEO. Videos can be optimized for engagement rather than accuracy. AI can produce confident answers even when the original question is based on a weak assumption.
Eventually I realized something uncomfortable.
I wasn't optimizing for truth.
I was optimizing for answers.
My first attempt
When large language models became part of my workflow, I started writing structured prompts.
Every new conversation began with the same instructions.
Define the role.
Explain the objective.
Specify the expected output.
Provide context.
The responses became noticeably better.
But every new conversation meant starting over.
My second attempt
To solve that, I built a project whose only job was generating better prompts.
Instead of writing the same instructions repeatedly, I described what I wanted, answered a few questions, and the project generated a complete prompt tailored to the task.
It improved consistency.
But I eventually realized I was solving the wrong problem.
I was optimizing the prompt.
Not the research process.
My third attempt
Today, my workflow looks very different.
I have specialized AI projects with custom instructions generated by another project I call my Prompt Engineer.
Its first job isn't to write prompts.
Its first job is to identify everything that's missing.
When I describe an idea, it doesn't immediately start answering.
It analyzes whether I provided enough context.
It identifies missing information.
It asks questions.
It suggests examples of the information I should provide.
Only after that does it generate the instructions for the specialized project.
The result isn't just a better prompt.
It's a better research environment.
The moment everything changed
One of these specialized projects helps me evaluate business ideas.
I described an application I'm building to track movies and TV shows.
My question was simple.
How can I monetize this product?
I expected recommendations about subscriptions, ads, or premium features.
Instead, before giving me any suggestions, the project expanded the scope of the research.
It started investigating licensing terms, copyright implications, metadata ownership, and whether I was even allowed to monetize the data source my application depended on.
I never asked about any of those things.
At first, I thought it had misunderstood my question.
It hadn't.
It had identified risks that directly affected whether my original question could even be answered responsibly.
That research completely changed my roadmap.
The MVP changed.
Version 1.0 changed.
Not because the AI gave me a clever monetization idea.
Because it challenged the assumptions behind my question.
What I consider noise
Today, I don't think bad research starts with wrong answers.
I think it starts much earlier.
Noise is:
Assuming instead of asking.
Storing information before verifying it.
Treating the first convincing explanation as the correct one.
Allowing unsupported assumptions to become part of future context.
Once that noise enters your system, every future decision becomes slightly less reliable.
What I actually want from AI
I don't want an AI that tells me I'm right.
I don't want an AI that answers as quickly as possible.
I want one that forces me to explain my reasoning before trying to solve my problem.
If I ask:
"Is Tesla the best investment?"
The answer I want isn't "yes" or "no."
The first question should be:
"Why do you believe Tesla is the best investment?"
That single question changes the entire investigation.
Because now we're no longer validating a conclusion.
We're examining the assumptions behind it.
The real upgrade
People often ask what changed in the way I use AI.
The biggest improvement wasn't switching models.
It wasn't writing better prompts.
It wasn't giving the AI more context.
The real upgrade was building a system that protects the quality of the research before generating an answer.
A system that asks questions.
A system that validates important claims.
A system that challenges assumptions.
A system that protects context from accumulating unverified information.
None of that guarantees perfect answers.
But it dramatically reduces the chances of making important decisions based on incomplete or incorrect premises.
And that's a much more valuable goal.
Because I've realized something over the last few months.
The quality of an answer depends far less on the intelligence of the model than on the quality of the thinking that shaped the question.
That's the kind of system I want to keep building.
Not AI that replaces thinking.
AI that improves it.
If this resonated with you, I'd love to hear how you're using AI in your own work.
Do you optimize for better answers, or have you started redesigning the way you think with AI?
You can simply reply to this email—I read every response.
If you know someone building with AI, feel free to share this article with them.
See you next week.
— Gilberto O.