The era of prompt engineering, I believe, is destined to be a transient one, with its existence spanning only a few years. The reason behind its emergence lies in the fact that the current generation of AI models can’t always capture our intentions and deliver the desired results.
— Me, 2023
Three years ago, I wrote that prompt engineering was probably temporary. The idea made sense to me then: prompts were a workaround for models that could not reliably infer what we meant. As the models got better, I assumed the workaround would disappear.
That prediction was right and wrong. Prompt Engineering back then mostly meant learning how to coax a chat model into a better answer: add examples, ask it to think step by step, specify the format, repeat your intent more carefully. That notion of Prompt Engineering is largely gone. But today Prompt Engineering is still alive and well because it’s transformed into a different concept. Today the important work is less about clever phrasing and more about making AI systems reliable when they have context, tools, memory, permissions, and real consequences.
The part I got right is that casual prompting got easier. You do not need to memorize a bag of incantations to get useful work out of a modern model. You can ask normal questions, use normal language, and expect the model to infer a lot more than it could in 2023. The old advice to say “think step by step” before every request now is rarely used.
Yet, that didn’t spell the end for Prompt Engineering. It changed to a new discipline. Once models became good enough to do more than answer questions, we started asking them to operate inside real workflows. Drafting an email is one thing. Choosing which email to send, deciding whether it needs approval, pulling context from the right thread, calling the right tool, and leaving an audit trail is something else.
If a chatbot misunderstands you, you get a bad answer. If an agent misunderstands you, it may mutate a database, message a customer, file the wrong ticket, or confidently skip a step that mattered. The more useful the model becomes, the more its instructions need to describe not just the task, but the boundaries around the task: what to do, what not to do, when to ask, when to stop, what context to trust, and what result counts as success.
Those instructions need to be systematically managed. That is engineering.
Part of today’s prompt engineering is context engineering. The model’s answer depends heavily on what you put in front of it: retrieved documents, prior messages, user preferences, account state, product rules, examples of good outputs, examples of bad outputs. Better models help, but they do not remove the need to decide what context matters. In many systems, the hard part is not getting the model to reason. It is making sure the model is reasoning over the right material.
It is also tool design. Once a model can call functions, browse files, query databases, send messages, or update records, the prompt has to teach it how to use those tools responsibly. Which tool should it prefer? What should it verify before taking action? What should it never do without confirmation? What should it do when two tools disagree? These are prompt engineering questions, even if they look more like product and systems questions than wordsmithing.
And then there are evals. If prompts are part of production behavior, they need the same discipline as other production artifacts. You need test cases. You need regression checks. You need to know when a model update changes behavior, when a prompt edit improves one path but breaks another, and when a workflow starts drifting from the outcome you intended. Prompt engineering becomes less about finding the perfect phrasing and more about building a feedback loop around the behavior you want.
This is why I no longer think prompt engineering is a phase. Every time models get better, we give them more context, more tools, more responsibility, and more ambiguous work. The model improves, and then the system around the model becomes more ambitious. Intent still has to be structured, boundaries still have to be represented, and production behavior still needs tests, monitoring, versioning, and rollback.
The 2023 version of prompt engineering probably was temporary. The folk art of magic words, clever phrasing, and ritualized instructions is fading as models improve. But the deeper discipline is not going away. It is becoming less like copywriting and more like systems engineering.
I was wrong because I thought better models would make prompts matter less. What actually happened is that better models made prompts worth taking seriously.