The Art of Talking to Machines

In high school I was handed a presentation on "non-Egyptian pyramids." I went home, typed exactly that phrase into a search engine, and within twenty minutes had fallen into a corner of the early internet that was extremely confident the Egyptian pyramids were built by aliens. My young brain had no conspiracy antibodies and no real history to push back with, so it just absorbed the whole thing like a sponge in a flood. I built an entire presentation on it. I later found out there are perfectly real pyramids outside Egypt, a fact I confirmed years afterward by standing in front of some in Mexico.

We have all done a version of this. Maybe not the aliens specifically. But the move of asking a powerful tool a slightly wrong question and confidently running with the answer is universal, and it stings to admit. The point that embarrassing memory hammers home is that the tool is only ever as good as your ability to talk to it. Truth bombs land hard sometimes.

The same thing is happening again

"Prompt engineering" is not a new idea, but what I am talking about is not memorising AI vocabulary like "few-shot learning," which just means showing examples, or "thought chaining," which just means asking for the steps. I teach at university now, and I regularly watch students ask AI to write Python while having no sense of what Python can actually do. The model hands back an unnecessarily elaborate solution to a simple problem, and they have no idea how to steer it somewhere better, because they do not know what better would even look like. It is genuinely painful to watch, in the way watching someone parallel park with their eyes closed is painful.

The skill is not the jargon. It is having a productive conversation, which matters more now than ever because of vibe coding, where you describe a problem in plain language and let the model produce the code. The term was coined by Andrej Karpathy in February 2025, and it moves the programmer from typing the code to guiding, testing and correcting it. The keyboard work shrinks. The conversation work does not. AI is not a mind reader. Not yet, anyway.

How to actually talk to it

These are the moves I have found most useful, all learned the dull way, by getting them wrong first.

When a response misses, do not start over, tell it precisely what is off. "That works, but make it handle empty fields properly." That is iterative refinement, and garbage in still produces garbage out.

When it has misunderstood the situation, hand it the missing piece. "This has to run on a phone with limited power, not a high-end machine." That is context augmentation, and context is most of the game.

When you are not satisfied, ask for options instead of one answer. "Give me three ways to do this, each with a different trade-off." That is alternative generation. The third one is occasionally the one.

When the goal is not obvious from the request, say the goal out loud. "I want this readable rather than maximally efficient, because junior developers will maintain it." That is goal articulation. Let it see a little of what you are actually optimising for.

When words are not landing, show it. "I need the output to look exactly like this." That is few-shot learning, and writing this section I am realising how much of my own time it quietly saves.

When you cannot tell why it did what it did, make it explain. "Why this approach and not the other one?" That is explainability prompting. The model occasionally needs to show its working, same as everyone.

And steer with comparatives rather than commands. "More concise." "Explain it like I am a beginner." That is directional feedback. You are nudging the wheel, not barking orders at it.

The half nobody sells you

Knowing when to stop is the same skill pointed in the other direction, and it is the part the courses skip because "stop using our subject" does not sell.

Step away when you genuinely need an expert. Some knowledge takes years and the stakes are too high to fake. Step away when the iterations stop improving things, because diminishing returns are a real and measurable cliff. Step away when verifying the AI's work would take longer than just doing the work, which I learned the hard and slow way. Step away when you want a genuinely new idea rather than a clever remix of existing ones, because remix is what the thing is built to do. Step away when you do not understand what it produced, because shipping something you cannot read is just posting future problems to yourself. And step away when the task needs live human collaboration, real-time adaptation, or the actual weight of an ethical choice, because the model does not feel that weight and will not pretend convincingly enough to matter.

That last category is the one to scale your caution against. Letting AI tidy the phrasing of an email needs roughly the oversight you would give a spellcheck. Letting it shape a decision you will have to defend out loud needs the kind of scrutiny you would give a stranger's directions before driving four hours on them.

The skill worth building

We had to learn how to search before search was useful. Now we have to learn how to converse before this is useful, and we are lucky that the internet is full of people figuring it out in public while we figure it out ourselves. I am not claiming I am good at this yet. I might be wrong about half of it. I think we are all working it out at the same time, which is its own small comfort.

The power was never in the technology. It is in whether you can have a real conversation with it and, just as importantly, recognise the moment to end one. So pick the next thing you hand to a model and, before you accept the answer, ask it the one question that tends to surface the most. "Why did you do it this way and not another?" Then decide, with the reasoning in front of you, whether to keep talking or close the laptop. By the way, I still occasionally Google things the wrong way. Old habits do not die. They just get quieter.