{"id":292,"date":"2026-10-08T15:47:06","date_gmt":"2026-10-08T15:47:06","guid":{"rendered":"https:\/\/prowad.pro\/is-prompt-engineering-still-worth-learning\/"},"modified":"2026-10-08T15:47:06","modified_gmt":"2026-10-08T15:47:06","slug":"is-prompt-engineering-still-worth-learning","status":"publish","type":"post","link":"https:\/\/prowad.pro\/fr\/is-prompt-engineering-still-worth-learning\/","title":{"rendered":"Is Prompt Engineering Still a Skill Worth Learning?"},"content":{"rendered":"

In March 2024, IEEE Spectrum ran a headline that still gets cited today: ‘AI Prompt Engineering Is Dead.’ It was based on real research, not speculation. VMware and Intel Labs researchers systematically tested hand-crafted prompts against automatically generated ones and found the automated prompts won. Two years later, the debate that headline started is still unresolved, and the honest answer is more useful than either side of it.<\/p>\n\n\n\n

The research behind the original claim<\/h2>\n\n\n\n

The IEEE Spectrum piece, by Dina Genkina, reported on work by Rick Battle and Teja Gollapudi testing prompt engineering strategies on grade-school math problems. Their systematic finding: prompts generated automatically by another AI model consistently outperformed prompts carefully hand-crafted by humans, including some of the specific wording tricks prompt-engineering guides recommended at the time. That is a real, testable result, not an opinion, and it undercut one specific claim: that hand-tuned wording is where the skill’s value lives.<\/p>\n\n\n\n

What replaced it in industry conversation: context engineering<\/h2>\n\n\n\n

Through 2025, a new term gained traction among AI practitioners, popularized in part by researcher Andrej Karpathy: context engineering, the practice of deciding what information, documents, examples, and tool access a model has available, rather than focusing narrowly on how a single instruction is phrased. Gartner reportedly advised enterprise AI leaders to shift investment accordingly, and Shopify CEO Tobi Lutke publicly said he preferred the term context engineering to describe what skilled AI practitioners actually do day to day. The shift in terminology reflects a shift in where practitioners say the leverage is: not the clever phrase, but what the model can see and use.<\/p>\n\n\n\n

A second, separate technical finding worth knowing<\/h2>\n\n\n\n

Chroma Research’s Context Rot study adds a distinct finding: model performance measurably degrades as input length grows, even well within a model’s advertised context window. This matters for the same reason the VMware finding does. It suggests skill is shifting away from wordsmithing a single prompt and toward managing what and how much you feed a model at once, a different and more structural kind of judgment call.<\/p>\n\n\n\n

The debate is real, not settled<\/h2>\n\n\n\n

It would be inaccurate to present this as resolved. Microsoft’s 31,000-person, 31-country workforce survey reportedly ranked Prompt Engineer near the bottom of new roles companies planned to add over the next 12 to 18 months, which suggests the narrow version of the skill has limited durable demand as a job title. But public commentary continues in both directions: Forbes contributor Bernard Marr argued in January 2026 that prompt engineering is no longer the most valuable skill, while fellow Forbes contributor Dan Fitzpatrick published a direct rebuttal in June 2026 titled Prompt Engineering Is Not Dead. Sam Altman also predicted back in 2022 that prompt engineering would not still be a discipline in five years, a prediction whose timeline is now close to being tested. We present this spread deliberately: the technical evidence points one direction, informed opinion remains split.<\/p>\n\n\n\n

What this means for teaching and training: our judgment<\/h2>\n\n\n\n

This section is our interpretation, not a settled fact. The narrow version of prompt engineering, memorizing specific magic phrases or formatting tricks, looks weaker as something to teach as a standalone skill; the VMware research and the Microsoft survey both point that way. What still looks durable, and worth building into any AI literacy curriculum, is the broader judgment underneath it: understanding what information a model needs to do a task well, recognizing when an answer is weak because the model lacked context rather than because the task is impossible, and iterating deliberately instead of hoping a clever wording will fix an underspecified request. That is closer to what context engineering describes, and it connects to the evaluation habits in our framework for evaluating an AI tool<\/a> and to what AI literacy actually means<\/a> as a durable competency rather than a tool-specific skill.<\/p>\n\n\n\n

Key takeaways<\/h2>\n\n\n\n