Here's my take: prompt engineering is not a real skill. It's a temporary workaround for interfaces that haven't matured yet, and the sooner we stop pretending it's a discipline on par with systems design or data modeling, the better off we'll all be. That doesn't mean prompts don't matter right now — they do. But building an identity or a career trajectory around the act of asking an LLM questions in a clever way is building on sand.
What We Actually Mean by "Prompt Engineering"
Let's be precise. When people say prompt engineering, they usually mean one of two things:
- Crafting a one-off prompt — writing a carefully worded instruction to get a better answer from ChatGPT, Claude, or Gemini for a specific task.
- Designing system prompts and pipelines — building structured, reusable prompt templates that feed into applications, agents, or automation workflows.
The first one is just using software. We don't call someone a "Google search engineer" because they know how to use quotation marks and site: operators. The second one is closer to real work — but it's software engineering with a natural-language configuration layer, not a standalone discipline.
The Steel-Man: Why People Think It Matters
I want to be fair here, because the other side has reasonable points.
Right now, LLMs are sensitive to phrasing. The difference between a vague prompt and a well-structured one — with context, constraints, examples, and a defined output format — can be the difference between useless output and genuinely useful results. That gap is real. I experience it daily. The prompt files I built for this very site are proof: a well-designed prompt saves me significant time and produces consistent, publishable content. Without the structure, I'd get generic filler.
There's also a knowledge component. Understanding how chain-of-thought reasoning works, knowing when to use few-shot examples versus zero-shot instructions, grasping token limits and context windows — these things matter when you're integrating LLMs into production systems. People who understand this stuff deliver better results than people who don't.
So the objection is: "If it produces measurably better outcomes, how is it not a skill?"
Why the Skill Has a Short Half-Life
Because the entire point of better AI is to make prompting less finicky. Every major model update reduces the gap between a naive prompt and an expert one. Consider the trajectory:
- GPT-3 (2020): You practically needed a PhD in prompt sorcery to get coherent output. Few-shot examples were mandatory. Formatting was a coin flip.
- GPT-4 / Claude 3 (2023-2024): Plain English instructions work well. You still benefit from structure, but the model meets you halfway.
- Current-generation models (2025): Models handle ambiguity better, ask clarifying questions, follow complex multi-step instructions without elaborate scaffolding.
The trend is unmistakable. The models are getting better at understanding intent from less precise input. The "skill" of prompt engineering is largely compensating for model limitations that are actively being eroded. Every time a model improves, the value of clever prompt phrasing drops.
An analogy that clicks for me
Early SQL databases required you to think carefully about join order because the query optimizer was primitive. People who could hand-tune query plans were valuable. Then optimizers got smarter, and that skill faded into a niche concern. The knowledge didn't become worthless overnight — but it stopped being a career differentiator. Prompt engineering is on the same path.
The Real Skill Is Underneath
Here's what I think people actually mean when they say someone is "good at prompting":
- They know what they want. They can decompose a vague goal into concrete, specific requirements. That's analysis and specification — a skill that predates LLMs by decades.
- They understand the domain. They can evaluate whether the output is correct, incomplete, or subtly wrong. That requires subject-matter expertise, not prompt expertise.
- They think in systems. They design reusable templates, chain outputs together, and build workflows. That's software engineering.
- They communicate clearly. They write precise, unambiguous instructions. That's technical writing.
Every one of these is a durable, transferable skill. None of them is unique to prompting. The prompt is just the interface through which these skills are expressed right now. Calling it "prompt engineering" elevates the medium over the message.
The Harm in Pretending Otherwise
This isn't just a semantic argument. Framing prompt engineering as a standalone skill causes real problems:
- It distorts hiring. I've seen job postings for "Prompt Engineers" offering six-figure salaries for a competency that will be largely automated within two to three years. Companies hiring for this title are buying a depreciating asset.
- It misleads junior developers. If you're early in your career and choosing where to invest your learning time, sinking hundreds of hours into prompt optimization instead of fundamentals like data structures, system design, or even clear technical writing is a bad bet.
- It creates a false ceiling. The people who are genuinely great at working with AI are great because they're great engineers or great thinkers — not because they memorized a library of prompt patterns.
Where I Land
Learn to write good prompts. Absolutely. The way you learn to write good commit messages, good bug reports, good documentation. It's a sub-skill within communication, not a discipline in itself. If your prompt engineering skill is built on top of strong domain knowledge, clear thinking, and software engineering fundamentals, it will serve you well no matter how the interfaces evolve. If it's built on memorizing that "act as a senior developer" gets better results than "help me code" — that's trivia, not expertise, and the next model update will make it irrelevant.
Prompt engineering as a real, durable skill? I don't buy it. But the skills that make someone good at prompting — clarity, specificity, systems thinking, domain expertise — those are as real as it gets. Invest there.
