Over the past two years I have watched prompt engineering appear and mature. The premise was straightforward: artificial intelligence responds to a clear instruction, and the better the instruction, the better the response. That premise is now quietly dissolving. The prompt is not disappearing, but its role is evolving.

The reason is worth examining, because it says something about the nature of understanding. In essence, a prompt is a compressed explanation of a problem. It asks the user to take everything relevant and squeeze it into a single, minimal instruction. Compression is efficient, but it is also lossy. The nuance lost in the compression is often the nuance that mattered.

Until recently, that level of summarisation was unavoidable. Early models were physically limited to taking in only a modest amount of text, or tokens, at once, so brevity was a technical necessity rather than a skill. That constraint is now largely gone. AI Agentic systems today run on context windows that have expanded from a few thousand tokens to hundreds of thousands, and in some cases beyond a million. A model can now hold an entire project, a year of correspondence, a set of contracts and the whole of a long conversation without losing a single thread.

When the constraint disappears, the optimal behaviour changes. There is no longer any advantage in compressing everything into one perfect instruction. It becomes far more effective to build understanding progressively — to say something, see what comes back, correct it, add a condition, challenge an assumption and go round again. Precision stops being a property of the request and becomes a property of the exchange.

This matters because these systems are, above all, context-dependent. Their output is a function of what they have been given to work with, and of how finely that material describes the problem at hand. A dialogue is simply a more efficient way of supplying that context than a single instruction, because it allows the relevant detail to surface progressively.

This is not a new idea. It is one of the oldest ideas we have. The Socratic dialogues were not a literary method; they were an argument about how knowledge works. Socrates rarely asserts; he asks, he follows an answer with a further question, exposes a contradiction, narrows a definition and asks again. The claim underlying the method is that truth is not transmitted from one mind to another like a parcel. It is uncovered, gradually, by two parties examining a problem together.

What is emerging in the interaction between humans and machines looks remarkably like that. Consider a manager preparing a complex project. The instinct of the prompt era was to write a detailed brief and ask for a plan. The dialogue approach begins differently. What is the actual objective? Which constraints are fixed and which are merely assumed? Which parts of this are genuinely understood and which are being guessed at? Each answer opens a further question. The project is decomposed into parts, and each part is examined on its own terms.

Only at the end does the user ask for a document. The plan, the summary, the risk register — these are no longer the starting point of the exercise but its residue. They record a conclusion that has already been reached rather than manufacturing one that has not.

This has a practical consequence that many organisations have yet to absorb. If the quality of the output depends on the quality of the exchange, then the scarce skill is not phrasing. It is questioning. The people who get the most from these systems are rarely the most technical. They are the ones who know what to ask, in what order, and when an answer is not yet good enough.

It would be a mistake to conclude that longer conversations are automatically better ones. A dialogue that wanders accumulates context without accumulating clarity, and context is not free: it must be read, held, questioned and ultimately reconciled by a human being. The discipline that made a good prompt valuable does not disappear. It simply moves from the summary of an idea to the structure of its examination.

Voice conversation with AI is consequently becoming the natural home for this, because speaking, for us, structures ideas faster, more simply and more naturally than typing ever did. The technology, as a result, is becoming sensory. Today, a person who cannot code but can hold a rigorous conversation with AI is well equipped for the work ahead. A person who can code but cannot think clearly, or express an idea precisely, will not be.

The prompt will survive, but as an opening move rather than a single instruction expected to resolve a problem while quietly carrying a set of unexamined assumptions. As our reliance on these systems increases, they will demand clearer and better-formed input — and the most reliable way to produce it will turn out to be the oldest one we have. Learning how to instruct machines will not be enough. We will have to learn, and grow comfortable with, reasoning out loud in front of technology that is listening critically — and to discover, in the exchange, what we actually think.