Prompt Engineering: Technical Skill or Scam?#

Prompt engineering was sold as the skill of the future. Courses, certifications, job titles, books, and even graduate programs sprang up around the idea that knowing how to talk to an AI was a differentiating technical competency. But was it really, or was it just a temporary patch for the limitations of language models?

The Origin: A Patch for Deficient Models#

To understand why prompt engineering existed, you need to remember what language models were like just a couple of years ago. GPT-3 and the first open models had limited natural language comprehension. They didn’t understand ambiguous instructions, lost context easily, and needed users to structure requests in very specific ways to produce useful results.

That’s where prompt engineering was born: a set of techniques —few-shot prompting, chain-of-thought, role prompting, delimiters, explicit output formats— designed to compensate for the models’ comprehension deficiencies. It wasn’t a skill about what you needed, but about how to ask a model that didn’t understand you well.

In other words, prompt engineering was the user interface that models lacked. It was the equivalent of memorizing terminal commands before graphical interfaces existed: necessary at the time, but destined to be absorbed by technological evolution.

How You Ask Is No Longer the Problem#

Today’s language models —Gemini 3, Claude Sonnet 4.5, GPT-5, Llama 4— are radically different. They understand natural language instructions without rigid formats. They can handle extended contexts, maintain coherence across long conversations, and most importantly, interpret the intent behind an imprecise request.

Do you need to tell Claude Sonnet to “act as an expert in X” to get a good answer? No. Do you need to structure your question with ### delimiters and few-shot examples for Gemini 3 Pro to understand what you want? No either. Models are now capable enough to understand you if you tell them what you need clearly, the way you would tell a competent colleague.

This doesn’t mean that how you communicate with AI is irrelevant. But the barrier is no longer technical. If you can express what you need clearly, the model will understand you. The problem is no longer the how, but the what.

The Exception: AI Development#

There’s an important nuance. For AI developers —those building processing pipelines, autonomous agents, or RAG systems— prompt engineering remains relevant as an optimization technique. In that context, structuring system prompts, controlling temperature, using few-shot for classification tasks, or defining JSON output formats are valid technical decisions that directly impact system performance.

But this is software engineering, not a “skill of the future” for everyday users. It’s like saying that knowing how to configure a web server is a universal skill because everyone uses the internet.

From Prompt Engineering to Context Engineering#

If prompt engineering was about how to ask the AI something, context engineering is about what you need and what information the AI should work with. This shift is fundamental.

Context engineering rests on two pillars:

1. Injected Context#

Modern AI agents don’t work solely with your message. They work with layers of context that you configure beforehand:

  • Context files: Define the general behavior, response style, constraints, and conventions the agent should always follow. They are configured globally or per project through files like AGENTS.md, CLAUDE.md, .cursorrules or .agent/rules (Antigravity), which the agent reads automatically to understand your code’s conventions, architecture, and technical decisions.

  • Agent skills: Specialized instructions for specific tasks that the agent loads when needed. For example, a skill for writing documentation, another for code review, another for generating tests. It’s not a magic prompt: it’s a workflow specification.

None of this requires knowing prompting techniques. It requires knowing how to define clear rules, document conventions, and structure your project’s knowledge.

2. Clear Specifications in the Prompt#

The second pillar is what you actually write in each interaction. And here’s the key: it’s not about how to write it (that’s prompt engineering), but what to write. The skills you need are:

  • Spec design: Knowing how to precisely describe what you want built. Functional requirements, constraints, edge cases, acceptance criteria. It’s the same thing you’d do when writing a well-crafted ticket or user story.

  • Solution design: Having a clear vision of how the solution should work. You don’t need to implement it, but you do need to know which components are involved, what data flows, and what outcome you expect.

  • Solution architecture: For complex tasks, knowing how to decompose the problem into manageable parts, define interfaces between components, and establish an execution order. This is engineering thinking, not prompting tricks.

In short: the skills that get you good results from AI are the same ones that make you a good professional. Clarity, precision, structured thinking, and the ability to decompose problems.

So, Was It a Scam?#

I wouldn’t say it was a deliberate scam, but it was a skill with an expiration date around which an entire industry was built. “Prompt engineering” courses that promise to transform your career in 2026 are selling obsolete knowledge. Prompting certifications are the current equivalent of a “web search” certification in 2005.

What you actually need to learn has nothing to do with formatting tricks. You need to learn to think clearly, to design solutions, to write specifications that leave no ambiguity. And if you work with AI agents, you need to learn to configure their context: rules, skills, and project documentation.

Prompt engineering is dead. Context engineering has replaced it. And the good news is that the skills you need to master it are the ones that have always defined a good professional.

References#

Anterior: Ingeniería de prompt: ¿Habilidad técnica o estafa?