“Prompt engineering is dead.” I keep hearing this. I disagree.
What actually happened is more interesting than an obituary. The craft didn’t vanish – it evolved into something structurally different, and the fingerprints are everywhere if you know where to look.
The Shift Everyone Noticed #
Agentic AI and Loop Engineering are the fashionable terms now. The conversation has moved from “write me a clever prompt” to “let the agent figure it out.” And honestly, it works. When you feed a model with relevant data that directly addresses your problem, the results speak for themselves. With RAGRAGRetrieval-Augmented Generation -- a technique that enables LLMs to retrieve and incorporate information from external data sources before generating a response. , information gets pulled from providers, files, vector databases – wherever your knowledge lives – and handed to the agent so it can actually grasp the task. Combine that with sub-agent orchestration and the enormous context windows we have today, and conversations start to feel smooth. Natural. Like talking to someone who already knows the context. Wikipedia 1 source 1 Retrieval-augmented generationWikipedia
You can even add another LLM layer that converts natural language into an engineered prompt before the real work begins. The user speaks casually; the translation layer sharpens it. The gap between intent and execution shrinks. And the prompting techniques themselves keep delivering. Chain of ThoughtChain of ThoughtA prompting technique where the model is guided to reason step by step before producing a final answer, improving accuracy on complex tasks. and Tree of ThoughtsTree of ThoughtsA prompting technique that explores multiple reasoning paths and evaluates them, allowing the model to deliberate and choose the strongest line of thinking. are deceptively simple. You guide the model to reason step by step, or you let it explore multiple reasoning paths and pick the strongest one. These methods can fundamentally change how an AI thinks – or at least encourage it to think in ways it otherwise wouldn’t. The results are quietly incredible. Wikipedia+1 2 sources 2 Large language modelWikipedia 3 Prompt engineeringWikipedia
So yes, the “just talk to it” approach works. I am not arguing against that.
The Cost Nobody Mentions #
But here is the complication. When you stop engineering your prompt, something has to hold the intent together. That something is the orchestration model – and it burns tool calls to do it. Every time a sub-agent needs to re-establish what you meant, every time context drifts and has to be recovered, you pay for it. Not in effort, but in tokens, in latency, in dollars. The convenience of natural conversation has a price tag, and it scales faster than people expect. When you run a team of LLMs against the same problem, that multiplier becomes visible fast.
There is a subtler problem too. Words have different meanings depending on context. Companies interpret terms one way, individuals another, and AI falls somewhere on that same spectrum. A word that feels obvious to you might mean something entirely different to a sub-agent three hops down the chain. If your orchestration doesn’t clearly communicate requirements and terms – or adjust tasks when meaning drifts – you get confident-sounding answers that are subtly wrong.
Where the Craft Actually Went #
This is where I part ways with the “prompt engineering is dead” crowd.
Prompt engineering is still a must-have skill today. Only the focus has shifted. It is not just about writing clever prompts anymore. It is about where those prompts live and how they are built. The soul files, the guardrails, the memories – the whole scaffolding around the model – that is where the real engineering happens now. That scaffolding helps keep intent, terms, and context clear across tools and sub-agents, which is exactly what the old hand-written prompt was trying to do for a single call. The harness gap between a bare model and a properly scaffolded one is the difference between a tool that drifts and a tool that holds. Wikipedia 1 source 3 Prompt engineeringWikipedia
I have been sharpening tools long enough to recognize the pattern. The work doesn’t disappear when the surface changes. It moves deeper. The person who only writes one clever prompt and hopes for the best gets predictable results. The person who builds the skeleton around the model – who decides what the agent knows, how it remembers, where it checks itself – gets something else entirely.
The scaffolding around the model is the new prompt engineering. And it never died. It just grew bones.