Open any social network today and you'll find someone with the ultimate prompt. The template "nobody tells you about", the list of magic phrases, the trick to save tokens, the fifty-page guide to asking for an email. It has become a market.
I'll be blunt, and I know not everyone agrees: most people don't have a prompt problem. They have a communication problem. And that problem wasn't born with AI. It was always there, in meetings, in Jira tickets, in campaign briefs. AI just pulled the rug off it.
The models are going the other way
There's an irony in this prompt market. With the early models, it made some sense to talk about technique: certain phrasings gave better results, repeating instructions helped, there were tricks that worked. The models were more fragile and you had to compensate.
Today's models go the opposite way. They understand natural language very well, pick up on intent, handle long context. And the very companies that build the models, in their own documentation, recommend something not very mystical at all: be clear, be direct, give context, explain what you want and why. Write the way you'd explain it to someone competent who just joined and doesn't know your project.
In other words, the bottleneck moved from syntax to clarity. The question stopped being "what's the right formula?" and became "do I know what I want, and can I say it?". And that second question is far more uncomfortable, because the answer is often no.
What communicating well with an AI looks like (and with anyone)
When I look at the requests that get a good result on the first try, they almost always have the same things, and none of them is a trick:
- The goal. What I want at the end, not the intermediate task.
- The context. Who it's for, why it matters, what has already been tried.
- The definition of done. How I'll know it's good. Length, format, level of detail, tone.
- The constraints. What's off limits, what's already decided, what's not up for discussion.
- An example, when it helps. Sometimes showing beats explaining.
Notice this isn't a prompt template. It's a list of things you need to have thought through before asking. If you can't write the definition of done, the problem isn't the AI. You still don't know what you want, and no magic prompt is going to solve that for you.
And the most interesting part: this list is exactly the same as a good brief, a good user story, a good request to a colleague.
Five years ago, the problem was exactly the same
Go back five years or so, before anyone talked about prompts. The ticket that reached engineering said "improve the checkout screen". Improve how? For whom? Measured by what? The engineer delivered what they understood, the business looked at it and said that wasn't it, and there went another week.
The meeting between business and tech where everyone agreed, and each person walked out with a different interpretation. The campaign brief that asked for "something impactful". The decision that was never written down and that, three months later, nobody remembered why it had been made.
I saw this many times, from both sides of the table, as an engineer and as the person asking. And in my view, communication noise has always been one of the things that most separated the companies that reach results quickly from the rest. Not the best technology, not the biggest team. It's the ability to get one person's intent to arrive intact in someone else's head.
The difference is that, before, that noise was absorbed by people. The senior engineer who came back with five questions before starting. The PM who translated the business request into something buildable. The alignment meeting nobody wanted, but that saved the project. There was a human cushion between a badly made request and the result.
AI removes the cushion, and amplifies everything
AI executes what you said, not what you meant. It may ask, but most of the time it fills the gaps with the most likely guess and delivers something with great confidence. Fast, well written and, if the request was vague, wrong in direction.
And here's the point I think few people are seeing: AI is an amplifier. Those who communicate well gain a speed that didn't exist before. Those who communicate poorly produce more wrong things, faster, and looking finished.
That applies to people and it applies to companies. A team that already had the habit of writing down decisions, defining what done means and giving context in requests is using AI to go much further. A team that lived on meetings without notes and one-line tickets is generating volume, not results. The gap between those two teams always existed. Now it has a lever.
Who is actually understanding the problem?
There's something that bothers me more and more, and it goes beyond the badly made request. Even with all the progress in AI, and with it more and more part of everyday work, new problems keep showing up. Every document generated by AI carries its understanding with it. And I've noticed engineering, product and business "benefiting" from that more and more: the spec comes out ready, the product document comes out ready, the plan comes out ready.
The question I keep asking myself is how far that benefit will go before it brings traps along with it. If we're gradually handing over even the understanding and the definition of what we want to AI, who is actually understanding the problem?
Even today we see companies putting effort into building before analyzing or validating. A conviction held just a little above the usual bar is already enough to steer an entire team. Now add to that equation a tool that tends to agree with whoever is asking and that turns any idea into a well-written, organized, convincing document. How would an AI not reinforce those convictions, leading whole teams, or worse, whole companies, in the wrong direction? Building something nobody asked for, changing what was good, or any other scenario. I wrote about validating before building in hypothesis before evidence.
I've caught myself many times asking business owners, colleagues, companies and peers: why are we doing this? And you'd be surprised how many times I was left without a clear answer. What scares me more and more is that AI is becoming the answer to everything, including for people who didn't have those answers or couldn't produce them.
And here is the other half of this post's argument. AI helps a lot those who already understand the problem and need to say it better. But for those who don't understand it yet, it delivers a document that looks like understanding. And a well-written document, inside a team, tends to be read as a decision already made.
Being fair to the other side
I don't want to fall into the opposite exaggeration. Technique exists and matters in some places. If you're building a product that calls a model thousands of times a day, prompt structure, token cost, output format and regression tests are real engineering, and they deserve care. I wrote a bit about that in AI in the product without hype.
But that's systems engineering, not the magic prompt sold for everyday use. And even there, what weighs most on the result, by far, is the clarity of the specification: what the system should do, for whom and within which limits. Communication, again.
What I'd do instead of collecting prompts
- Train writing, not formulas. Writing is thinking. People who write clearly think clearly, and that holds for a request to an AI, a product document or a message to the team.
- Write the definition of done before asking. If it doesn't come out, stop and think about what you want. That's the real work.
- Ask "why are we doing this?" before asking for the document. If nobody on the team can answer without opening the chat, the AI won't answer it for you. It will just fill the gap with something convincing.
- Read the wrong result as a mirror. When the AI delivers something off from what you wanted, first ask what was ambiguous in your request. Most of the time, the answer is right there.
- Take it to the team. Written decisions, context in requests, definition of done on tickets. Everything that improves communication between people improves AI use, for free.
- Be suspicious of anyone selling shortcuts. If the promise is results without you having to think about what you want, it doesn't hold up. It never did.
Summary
The perfect prompt is good communication, and it always was. Models keep getting better at understanding natural language, which makes it ever clearer that the bottleneck is the clarity of whoever is asking. That bottleneck separated companies long before AI: in the vague ticket, the generic brief, the meeting with no decision. The difference is that now there's no human cushion absorbing the noise, and there is a tool ready to turn an unanswered conviction into a convincing document. Understand before you ask. AI amplifies what you already are. If you communicate well, it takes you far. If you don't, it just helps you get things wrong faster.
Murilo Anzini
Hands-on writing on engineering, product, data and marketing for people who want to understand, and build, the whole journey of a product.
In times when everything comes from AI, I like to write because it reinforces my learning and it's my moment of personal growth, and sharing it with you is a pleasure. The ideas, opinions and stories are mine; AI helps me structure, review and translate, because I'm only human.