How to Talk to a Machine

By Ken Ho

Anyone who has spent time with a small child knows the feeling. A casual instruction — “make something nice” — yields chaos. Blocks fly. Nothing resembles the intended outcome. The adult sighs and tries again, this time with more care: place the red block on the blue one, precisely here, like this. Slowly, a tower takes shape.

It is a scene played out in living rooms around the world. It is also, perhaps surprisingly, a useful guide to three of the most consequential skills of the present decade — skills that go by the names prompt engineering, context engineering, and harness engineering. The jargon is unfortunate, because the ideas behind it are straightforward. Each one answers a simple question: how do you get a powerful but unpredictable system to do what you actually want?

The answer, it turns out, looks a lot like teaching a toddler to build a tower.

Prompt engineering: say exactly what you mean.

The first discipline is the one most people have encountered, even if they have never called it that. Prompt engineering is the art of writing clear, specific instructions for an AI system.

Consider the toddler. “Make something pretty” is an invitation to chaos. The block might end up anywhere — or sailing towards the cat. But “put the red block directly on top of the blue block” leaves little room for misinterpretation. The child knows precisely what to do.

AI works the same way. “Write something about technology” is the digital equivalent of handing over a bucket of blocks and hoping for the best. The result will be something, certainly, but unlikely to be what was wanted. Precision changes the equation. “Summarise this ten-page report into three bullet points, each under thirty words, focusing on financial risk” tells the system the format, the length, and the subject. Specificity does not guarantee perfection, but it dramatically narrows the gap between intention and output.

This is not a new insight. Engineers have always known that garbage in produces garbage out. What is new is that the input language is now natural English — and with that convenience comes a temptation to be lazy, as if the machine should simply understand. It does not. It responds to structure. Prompt engineering is the discipline of providing that structure through the instruction itself.

Context engineering: set the scene before asking for results.

No sensible parent asks a child to build a tower on a shaggy carpet. The blocks need a hard, flat surface — a stable foundation that makes success possible. The environment matters before the task even begins.

In AI, this principle has acquired a name of its own: context engineering. If prompt engineering is about what you ask, context engineering is about everything surrounding the ask — the examples, the background, the constraints that frame the request before a single word of the actual instruction is given.

Providing a relevant example — “here is the sort of output I am looking for” — gives the system a template. So does background information: the sector, the audience, the tone. A prompt that begins “you are a regulatory compliance officer reviewing a banking document” produces categorically different results from one that simply asks “is this document compliant?”

The difference matters. Context reduces the cognitive burden on both sides. The system spends less effort interpreting the request and more effort executing it. The human spends less time rephrasing and more time refining. Think of it as clearing the table before the toddler begins: the blocks are the same, the child is the same, but the chances of a standing tower improve dramatically.

Harness engineering: build guardrails, not cages.

Watch a careful parent with a building toddler and you will notice something subtle. They form a V shape with their legs, creating a soft boundary around the child. If the tower wobbles, the parent is there. The child still builds freely, but failure is contained.

The digital equivalent is called harness engineering — the practice of designing safety measures around an AI system without smothering its usefulness.

This means filtering inputs to avoid overwhelming the model with irrelevant data. Setting length limits and format constraints. Validating outputs before they reach an end user, so that nonsense is caught before it causes harm. These are not restrictions on capability; they are the conditions under which reliable capability becomes possible.

The most effective systems anticipate what could go wrong and quietly prevent it. They do not eliminate risk; they contain it. The result is not a system that never fails, but one whose failures are smaller, more predictable, and easier to correct — just as the parent’s legs do not build the tower, but ensure that when it topples, nothing breaks.

Three disciplines, one principle.

Prompt engineering, context engineering, and harness engineering. Three names for what is, at heart, a single idea: communicating with care. The first shapes the instruction. The second shapes the surroundings. The third shapes the safety net. Together, they turn a system that might do anything into one that does something useful.

None of this requires a technical background. The principles are the same ones that govern any successful collaboration between humans — be clear about what you want, provide enough background for the other party to understand, and check the work before it goes out the door. The only difference is that the other party is now a machine.

As AI tools become embedded in more workplaces — from drafting legal documents to summarising medical records — the gap between those who can direct these systems effectively and those who cannot will widen. Not because the technology is especially complicated, but because it rewards clarity in exactly the way that most human communication does not require.

The toddler, at least, is forgiving. The machine is not. It will do precisely what is asked — which is why learning to ask precisely matters more than ever.