Prompt Engineering Is a Real Skill. Stop Pretending It Isn’t.
And stop acting like you need a CS degree to learn it.
My first “serious” prompt was embarrassing.
I asked AI to “Write me a marketing email.” And received some boring generic response, which I deemed absolutely useless. I had assumed that there was something wrong with the product, but no, it was simply my incompetence and lack of knowledge on how to use it better.
Now, after eight months of working with the same tools and models, I get responses that my colleagues look at me like: “How did you manage to do that?”
I have taken prompt engineering seriously and studied it, as with any skill. Not as something entertaining or cheating. As an actual practice with rules and techniques.
And honestly, that’s where things get frustrating. A lot of people online would rather mock someone for trying to learn prompting than help them get better at it.
The Gatekeeping Is Real — and It’s Annoying
There’s one type of individual that comes out when people start talking about prompt engineering in earnest. They always seem to say something along the lines of, “It’s just asking questions; there’s no need for a course on that.” Or the classic one, “Engineers construct the models; prompt engineers only speak to them.”
That logic has never made much sense to me. Nobody ever doubts the craftsmanship involved in writing dialogue in a screenplay or choosing the right words as a copywriter. Every tool requires some form of craft. The fact that the user interface uses natural language does not diminish the need for skill.
Those who engage in this type of gatekeeping usually fall into one of two camps. One, developers who feel under attack since their skill suddenly means little, and secondly, contrarians who have never actually taken the time to properly test a language model. They both know nothing of what it takes.
And here’s what it looks like to do prompt engineering.
What Prompt Engineering Actually Is
But let’s be clear, because people like to abuse terms.
It doesn’t involve using manipulation techniques to make AI do something it shouldn’t be able to do. It doesn’t involve finding ways to bypass the model’s limitations. It involves crafting prompts clear enough for a model to carry out the desired actions.
This all may seem easy. But it’s not.
Models are trained on patterns. They don’t think — they predict. This means that whenever a model fails to generate a satisfactory response, the issue is hardly ever that it failed to perform. Instead, there was something wrong with the prompt itself, which made it possible for the model to go down the wrong route in its predictions.
A great prompt engineer understands these basic principles. And knows that it comes down to asking what kind of completion is prompted by this prompt. And then writing until this particular kind of completion becomes more probable.
This involves context, tone, style, specificity — sometimes, even knowledge of the model’s limitations.
The Four Things That Actually Improve AI Outputs
Forget the 47-step prompting frameworks you see on Threads. Here’s what works in practice. Once you understand these four things, working with AI starts feeling far less random.
1. Set the context before the request.
Many people start with their question and end there. “Write a product description for my SaaS tool” may get you a decent product description. But “You are a conversion copywriter specializing in B2B SaaS. Your job is to write product descriptions that help skeptical buyers quickly grasp the value.” Followed by the same question, you will get a product description you can actually use.
This isn’t clear to the model until you explicitly state it. Tell it who it is and what it’s supposed to do before making something. Models will respond better when they understand the role, context, and objective before generating anything.
2. Constraints usually improve the result
It’s no surprise that vague prompts lead to vague outputs, but here’s the trick — Adding constraints doesn’t mean lower quality; it means higher quality.
“Write me a LinkedIn post about leadership” is too open-ended. “Write me a 150-word LinkedIn post about leadership targeted at mid-level managers who often feel overlooked, with a tone that is direct and frank, not motivational” is a good brief. The more constraints you include, the fewer wrong paths the model will take to the goal.
3. Show examples instead of overexplaining
This is one of the most overlooked prompting habits. Instead of spending five paragraphs explaining the style you want, just provide an example.
Paste a previous output you liked. Or a piece of writing with the structure or tone you want the model to follow.
AI models are surprisingly good at recognizing patterns once you give them something concrete to imitate.
4. Treat prompting like an iterative process
Many people think bad output means a poor prompt. Experienced users usually do the opposite. They adjust and continue.
How was the result too generic? Too stiff? Off topic? Didn’t emphasize the key points? Each bad output gives you information for the next version of the prompt.
The process isn’t “write prompt → output → accept/reject,” but “write prompt → review output → edit prompt → repeat.”
That’s why prompting works better as an iterative process rather than a one-shot transaction.
The Embarrassing Truth About “Just Talking to It”
What a lot of people fail to realize is that natural language can actually be a messier interface than traditional software.
In the case of a command-line program, everything is clearly defined: the user knows how the program will react to input, which inputs will be considered by the program and which will be ignored, and how each option works. In the case of a language model, the user is operating a program that contains all the ambiguities of the English language. There is a bleed from context. Previous statements can influence subsequent statements. One word can change the tone of an entire reply.
People who are good at prompting are constantly managing tone, context, intent, structure, and clarity all at the same time. They’re guiding a complicated system using language instead of code.
And unlike programming, there’s usually no syntax error telling you what went wrong.
You just get a bad answer and have to figure out why.
Why This Matters Beyond Productivity Hacks
I think the conversation around prompting is bigger than simple productivity advice. Calling it “just a productivity hack” doesn’t really capture what’s happening right now.
We are at the start of a time when our interaction with AI systems becomes the mediator of vast amounts of cognition. Everything from research to writing to analysis to design to planning happens through AI systems today. The ones who can speak their language clearly will develop a compounding advantage over others.
In fact, the people with both technical expertise and strong prompting skills will probably have the biggest advantage. AI works best as a multiplier, not a replacement.
But the ability to dismiss it as an option is fast vanishing. The more powerful these systems get, the broader the set of use cases and the wider the skill gap grows. The people ridiculing the idea of prompt engineering in 2023 were the ones ridiculing anyone with good Excel formula knowledge in 1995. Maybe I’d check in with them now.
A Few Things Worth Trying This Week
If you want to actually build this skill, here’s a shortcut:
Pick one use case and go deep. Don’t try to optimize your entire workflow at once. Take one thing you do repeatedly — summarizing research, drafting responses, writing code comments — and spend a week crafting, testing, and refining a prompt for just that.
Keep a prompt log. Seriously. A plain text file where you save prompts that worked and prompts that failed, with notes on why. The pattern recognition you develop from reviewing your own history is worth more than any framework.
Read outputs. Train yourself to notice when something is almost right but not quite. That gap — between “good enough” and “exactly right” — is where most of the craft lives.
Study the model’s weak spots. Every model has tendencies: things it does when it’s uncertain, ways it pads low-confidence responses, structures it defaults to when your intent is ambiguous. Once you start noticing those patterns, you get much better at correcting problems before they even happen.
The Skill Isn’t Going Anywhere
There’s a version of this argument that says prompting will become trash as models get smarter. Maybe. But so far, the opposite seems to be happening. Better models usually make skilled prompting more useful, not less. As the systems become more capable, the gap between good prompting and bad prompting becomes easier to notice.
What does seem to be disappearing is the idea that prompting is just random luck — that some people magically “have it” while others don’t.
That was never really true.
Like most skills, people improve when they study it seriously, experiment consistently, and pay attention to what actually works.
And the gatekeeping will probably continue. People protect their status by narrowing the definition of what counts as real skill. It’s a very human thing to do.
But you don’t really need permission to get good at this. The work is there. The improvement is real. And the outputs speak for themselves.
Want to go deeper? The best learning happens by doing — pick a model, pick a use case, and start iterating. No course required.
Prompt Engineering Is a Real Skill. Stop Pretending It Isn’t. was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
