Has Prompt Engineering Died? From Phrasing to Systems Engineering

In June 2025, Andrej Karpathy (one of the most prominent figures in AI) posted a tweet that quietly ended a three-year debate. He said, in essence, that people associate "prompts" with short task descriptions you give a model in everyday use, while in every serious AI application, what actually matters is "context engineering": the delicate art of filling the context window with just the right information for the next step. Within hours, Shopify's CEO amplified the message, and by July Gartner published a headline that would have seemed provocative a year earlier: "Context engineering is in, and prompt engineering is out."

So has prompt engineering really died? The precise answer: no, but it is no longer the whole story. It shifted from a standalone craft into a single component within a larger, more complex system.

What Exactly Changed?

In 2023, crafting the perfect prompt felt like magic. You write "you are an expert engineer, think step by step," and the output genuinely improves. By 2025 it became a marketable skill. But as models matured, these "tricks" began to give diminishing returns. The reason is fundamental: modern models have become too smart to need incantations of phrasing, while the real problem moved elsewhere — to the information available to the model at the moment of decision, not to how you ask it.

The difference between the two concepts can be summed up like this: prompt engineering optimizes the "question," while context engineering optimizes the "conditions under which the question is answered." The first is tactical, the second architectural.

Why Is Clever Phrasing No Longer Enough?

The answer lies in the nature of agents. A single prompt works for a single interaction, but an agent executes a chain of steps, each passing its state to the next. Here a failure mode appears that has nothing to do with phrasing at all: an agent works flawlessly in one step, then its successor in the chain receives an incomplete state and begins to "hallucinate" information that was originally there but lost in the handoff between steps. The model did not fail; the context handoff failed. As Philipp Schmid of Google DeepMind put it bluntly: "Most agent failures today are not model failures, but context failures."

This is why hours of prompt tuning are useless for fixing what was never a prompt problem to begin with. The agent does not fail at its steps, but between them — because the wrong information is present at each node.

Context Engineering: The Five Layers

Context engineering is not a replacement for the prompt, but a broader framework that contains it. The 2026 literature describes it as a stack of five layers: the instruction layer (including the system prompt, where the craft of phrasing stays alive), the retrieval layer (fetching the right documents via systems like RAG), the memory layer (preserving context across long conversations), the tools layer (what the model can call), and the state layer (what is preserved between agent steps). A well-crafted system prompt is still essential, but it has become one input in a much larger system: the text layer in a five-layer stack.

Where Does the MCP Protocol Come In?

This shift is accelerated by the spread of the Model Context Protocol (MCP), whose public servers crossed ten thousand by late 2025. What MCP does architecturally is "externalize" context sources, so the agent can fetch what it needs when it needs it rather than receiving everything pre-loaded. The most accurate analogy: the difference between a library and a briefing document. The agent goes to the shelf when it needs a specific book, instead of carrying the whole library with it at every step.

So Has the Prompt Skill Vanished?

Here balance is needed. The claim that prompt engineering "completely died" is an exaggeration more popular on social media than an engineering truth. Crafting a clear, precise system prompt is still a core skill, and it is one of the five layers, not the least important of them. What actually ended is the illusion that the "magic word" alone solves everything. The significance of this shift is in the numbers: a 2026 survey found that 82% of IT and data leaders agree the prompt alone is no longer sufficient for production, and 95% of data teams plan to invest in context engineering capabilities.

What Does This Mean for Your Career?

If you built your skill on "prompt tricks" alone, the troubling news is that these tricks do not "compound": you learn a trick that works for one case, then forget it after months, with no system being built. Context engineering, however, is architectural and compounds: you build a retrieval system, design a memory, organize tools, and each project becomes a brick in reusable expertise. The practical advice to start: use the persistent context features available today (like Projects in Claude), build a context pack for each work domain, then study how the layers are assembled into a real agent system.

The irony is that this shift repeats a historical pattern: just as web design evolved into software engineering, and copywriting into growth marketing, the original craft did not vanish but became a tool within a more complex stack. Prompt engineering did not die; it graduated into systems engineering. And those who understand this difference today build a skill that lasts, not a trick that withers with next month's model.