← AI for Educators Daily with Dan Fitzpatrick
AI for Educators Daily with Dan Fitzpatrick
Is 'prompt engineering' still vital for teachers?
00:11:19
Loading summary
Transcript1 lines
- [00:00]
A
If this episode makes you think, please let us know in the comments and support us by subscribing and leaving a review. Thank you. Today, we are exploring an article I wrote for Forbes this week, simply titled Prompt Engineering Isn't Dead, but the Caricature Is. It's a piece where I tried to cut through some of the noise about a topic that's often talked about but rarely deeply understood. Now, you might have seen headlines confidently declaring the death of prompt engineering. It's become a bit of a cliche, hasn't it? The idea that this skill of optimizing instructions for AI models like ChatGPT or Gemini is just gone. Finished. Obsolete. And to be fair, I can see why people might think that the language around it has certainly changed. The article mentions that in June 2025 ch just a year ago, the AI researcher Andre Karpathy popularized a much tidier term, calling it context engineering. He described it as this delicate art of filling a model's context window with exactly the right information at each step. And then a month later, you had Gartner, that big research and advisory company, telling its clients that context engineering is in and prompt engineering is out. It all sounds very definitive, doesn't it? But here's the thing, and this is what I was arguing in the piece. If you're celebrating the funeral of prompt engineering, then I think you might never have really understood what it was in the first place. Someone somewhere did coin the phrase prompt engineering, and I'll admit it probably sounded a bit pretentious, a bit overly technical. But when you strip that label away, what you're left with is actually a very old, very, very human skill. It's simply the refinement of language to express an idea clearly enough that it can be acted upon. Think about that for a moment, making your intent legible to someone or something else. Isn't that what we do every single day? It's a skill as old as teaching, as old as managing people, as old as raising children, frankly. And that skill, that fundamental ability to communicate clearly, that isn't going anywhere. If anything, what this whole debate is showing us is how easily we can get thoroughly confused by the label, by the word we use for the thing it actually points at. Now, the obituaries for prompt engineering do get one part right. The AI models themselves have changed, and they've changed substantially since 2022. Early systems, when they first came out, rewarded a kind of incantation. You'd type in these magic phrases or use specific roleplay tricks, maybe even try an odd threat or bribe just to squeeze out a slightly better response from the chatbot. That kind of fiddling with magic words that barely matters now. The article makes the point that what really moves the needle today is how you structure the whole task, the entire interaction, rather than just polishing one perfect sentence. Tech entrepreneur Andrew Eng really made this plain when he spoke at Sequoia Capital's AI Ascent event back in 2024. He shared some fascinating results from the human evil coding benchmark. So imagine they gave GPT 3.5, which is an earlier model, a single prompt. It solved roughly 48% of the problems. The stronger model, GPT4, did better, solving about 67%. But here's where it gets interesting. Ng's team took the weaker model, GPT 3.5, and wrapped it in an iterative loop. They let it draft its answer, then check its own work, then revise across several passes. And what happened? Its accuracy climbed to about 95%, beating GPT4's single attempt outright. What's the lesson there for us? For educators, it's that we need to stop just thinking about polishing one perfect sentence and instead start designing the systems or the environment that the AI model works inside. And that, in essence, is what context engineering truly means. So yes, the article concedes, fiddling with those magic words is indeed a dead end. The critics are right about that part. But it's the conclusion they've drawn, that the skill itself is dead that's wrong. We're not teaching students to outsmart machines with clever tricks. We're teaching them to outthink them by designing better processes. Processes. This is classic human in the loop thinking, isn't it? Humans remain the decision makers, setting up the conditions for AI to excel and then reviewing what it produces. British programmer Simon Willison, who actually argued for prompt engineering as a serious craft from the beginning, really named the real problem here. He admitted that most people's working definition of it had shrunk to a laughably pretentious term for typing things into a chatbot. And he's right, isn't he? The actual discipline, the careful, thoughtful work of communicating intent, just disappeared under two years of viral threads about jailbreaks and secret phrases. We ended up killing off the caricature of the skill, and then told ourselves we'd killed the craft itself. But look at what context engineering actually demands of you. You have to decide what a model needs to know, and in what order and under what constraints. And, crucially, what it must never do. Is that a new skill? Not at all. It's an old one, simply scaled up for a new tool. It's that same refinement of language to express an idea clearly enough that it can be acted upon. This is a human skill before it is ever a technical one, and it behaves the exact same way. When you're interacting with people, think about how you talk to a colleague who knows you really well. You probably say less or you phrase things differently because you have shared context, don't you? Now, think about how you brief a junior member of staff who's just starting out or who's growing more capable. You adjust your communication, you say more, or you break it down differently. We do exactly that with AI. As it improves and as our understanding of it deepens, the specific words we reach for might change, but the fundamental premise underneath them never does. It's always about knowing what you want to achieve and then saying it clearly enough to be understood. I spend most of my work in life with educators trying to help them get something meaningful out of these AI tools, rather than something that just merely looks impressive. And I see this pattern show up constantly. The quality of the result depends far less on the specific AI model being used and far, far more on what the educator asks and how precisely they ask it. The teacher who knows exactly what she's trying to achieve, who has a clear educational purpose in mind and can put that into clear, articulate words, gets the better outcome every single time. That's purpose over technology in action, right? Start with the why, not the how. So why does any of this matter for what schools teach? Well, the OECD's Digital Education Outlook 2026, which came out in January, contained a warning that I think is really easy to skim past. It basically said that when students hand tasks over to general purpose AI without a clear learning purpose, their output often improves, but their learning doesn't. And those gains, they tend to vanish the moment the tool is taken away. The report's conclusion is striking. Critical thinking and higher order metacognitive skills matter more, not less in this new AI era. And what's notably missing from that conclusion? Prompt syntax, those magic words. The skills that endure are thinking and judgment. This is about design and learning that cannot be faked because it demands depth, care and imagination from the student, not just an AI generated product. The real value, as I often say, is not in what the machine produces, but in how the student responds to it, how they engage with it. The European Commission and OECD's Joint AI Literacy Framework, which is due out in full this year, goes even Further, it builds AI literacy around four key capabilities. Engaging with AI, creating with it, managing it, and designing it. And it defines managing AI as delegating tasks with clear rules and proper human oversight. Think about that. Delegating clearly with the judgment to know when and how to step in, when to critically review. That's not some specialist technical skill that requires hours of coding. It's fundamentally about communication and critical thought. The framework's own authors say they specifically set out to name skills that will remain durable as the technology continues to shift and evolve. This is evolution, not revolution in practice, for how we approach AI literacy in schools. It's about teaching students not to just use tools, but to think with them, understanding their limitations, and managing conversations with precision. I've watched schools pour a tremendous amount of effort into mastering whichever specific AI tool sat in front of them that term, only to find both the tool and the technique were dated or even obsolete within a year. And that's disheartening for teachers, isn't it? It feels like chasing a constantly moving target. The schools, I think, are really getting this right, aren't chasing the tool at all. They're doing something much more fundamental and much more enduring. They're teaching young people to think clearly and to say what they mean. They're focusing on the process and productive struggle of learning, where the real growth happens. So, no, I wouldn't mourn prompt engineering, and I certainly wouldn't be crowning context engineering in its place as the next big thing. They are essentially the same human capacity, just wearing different clothes for a different season. The term itself will be retired and renamed again. I'm absolutely sure of it. For anyone who teaches. For any educator listening, the real move here is to stop tracking the buzzword of the moment, stop chasing the latest label. Instead, teach what sits beneath it all. Ask yourself. And encourage your students to ask themselves, what am I really trying to achieve here? And then have I said it clearly enough that anyone or anything could possibly act on it? That fundamental question worked long before AI ever arrived on the scene, and it will still be working when the next buzzword has its funeral. That's all for today. Thanks for listening.