By Debleena Majumdar & Arjo Basu
“If you don’t know where you are going, any road will get you there.”
That line from Alice’s Adventures in Wonderland reads today less like whimsy and more like a warning label for artificial intelligence.
It sounds comforting. Until the road starts talking back.
We now live with machines that answer instantly. They are fluent. Polished. Untroubled by doubt. Ask a question and you get a response. Ask again and you get another—equally confident, slightly different, just as sure of itself.
Somewhere between the question and the answer sits a small, easily ignored thing. A sentence. A paragraph. Sometimes just a few words typed into a box.
We call it a prompt.
It feels incidental. Almost disposable. But this is where intent either sharpens—or quietly dissolves. The machine does not pause to ask what you really meant. It assumes. And assumption, at scale, is never neutral.
The small thing doing all the work
Between human intent and machine output sits something so ordinary it rarely attracts attention: the prompt.
It looks trivial. It isn’t. This small act is doing all the real work. It carries the entire burden of translation, from what a human half-means to what a machine will confidently do. On paper, the definition is clean enough. A prompt is the input we provide to a Large Language Model (LLM) to generate an output. True. And also beside the point.
Because what matters is not what a prompt is. It’s what a prompt does.
A prompt compresses intent. It takes something vague—an expectation, a curiosity, a demand—and turns it into a signal that a probabilistic system can act on.
When that signal is clear, things mostly behave. When it isn’t, the system doesn’t slow down or ask for clarification. It begins to assume. It fills in the gaps. And it does so fluently, confidently, and without hesitation. At scale, that behaviour stops being harmless.
Why everything keeps sounding the same
Spend enough time prompting and a pattern emerges.
The same prompt gives different answers on different models. The same model gives different answers on different runs. And over time, the answers begin to blur.
They sound balanced. Reasonable. Agreeable. They avoid offence. They smooth over disagreement. They leave very little behind. The sharp edges disappear. The odd insight gets ironed out. What remains is a polished middle.
This is not an accident. It is the system doing exactly what it is designed to do.
When intent is underspecified, language models converge on what is most statistically likely to be acceptable. Not the most accurate. Not the most interesting. The most defensible. The least objectionable.
This is the statistical median. It masquerades as wisdom.
The median is safe. The median avoids strong commitments. The median sounds like it knows what it’s doing. But the median is also where thinking quietly goes to die.
For those who spend real time with these systems, this is where frustration sets in. And with it, a tempting idea: that somewhere there exists a perfect prompt. A secret syntax. A phrase that finally unlocks depth.
There isn’t. What exists instead is discipline. And a clearer understanding of what you are actually dealing with.
What the machine actually is
Despite the name, a LLM does not know things. It does not retrieve facts the way a database does. It does not reason the way humans do.
What it does instead is predict the next most likely word in a sequence, based on patterns learned from vast quantities of text. That is where the fluency comes from. It is also why the system can sound brilliantly persuasive while being hopelessly wrong.
When intent is underspecified, the model substitutes an average assumption drawn from its training data. Statistically speaking, it reaches for the middle of the distribution.
Unless something pushes it away from that centre, that is where most prompts end up.
Making intent legible
A good prompt does not extract intelligence from a machine. It makes intent legible to one. At the very least, it does three things.
It clarifies purpose.
It constrains what counts as an acceptable answer.
And it gives the system a way to recognise when it has done a decent job.
None of this is exotic. It is familiar territory. Because this is how good writing works as well. Someone speaks. Something happens. It happens somewhere. For a reason.
In prompting, these fundamentals reappear—quietly—in three layers.
So what is the prompt trying to do, anyway?
The first layer is structure. It answers a basic question the system faces every time.
What am I doing here?
Four signals matter more than the rest.
Role. Who is the model supposed to be? A macroeconomist, a product manager, a regulatory analyst, a literary editor. Each role activates a different slice of its training.
Task. Analysis is not synthesis. Critique is not explanation. Weak prompts blur these together. Strong ones do not.
Audience. Writing for a general reader and writing for a board are different acts, even when the topic is identical. When the audience is unspecified, the model defaults to the widest possible one.
Output. Length, structure, format. These are not cosmetic choices. They shape the thinking itself.
Leave any of these unstated, and the median steps in, happy to help.
Escaping the median
Structure gets you out of confusion. It does not get you out of blandness.
For that, prompts need principles.
These are the standards by which the answer will be judged: Rigor. Scepticism. Originality. Decisiveness.
Principles also rule things out: excessive hedging, false balance, motivational fluff.
This is where many prompts fail without drama. They never specify what a good answer looks like. The model is left guessing. Each run becomes a fresh attempt to sound pleasing.
Once principles are explicit, something changes. Variation drops. Focus sharpens. The system’s wandering is now within tighter bounds. It has a standard to aim for.
Don’t make it think all at once
Even with structure and principles, complex prompts often stumble for a simpler reason. They ask the model to do everything in one go.
Language models behave better when reasoning is staged.
First identify assumptions.
Then examine options.
Then arrive at conclusions.
Tell the model how to think before asking it what to say, and the output becomes more consistent, more inspectable, more useful.
Not smarter. Just less sloppy.
Where prompts are actually learned
In practice, prompting isn’t learned in theory. It is learned under pressure.
Someone sends you a long document and asks for a summary. A boss wants a slide rewritten before a call. A piece of code throws an error and you just want it fixed.
So you type something in. Anything. You see what comes back. You adjust.
If the answer works, you move on. If it doesn’t, you try again.
Over time, a kind of intuition forms. You learn what passes. What gets questioned. What causes trouble. You learn how cautious an answer needs to sound. How confident is too confident. Which words are better avoided altogether.
None of this is written down. It settles into muscle memory. That is how “good enough” takes shape—not as a standard, but as a habit.
The tools adapt quickly. Some push you to be clearer, more deliberate. Others reward speed and fluency and leave the rest to chance.
Either way, prompting does not grow out of theory. It grows out of use.
Beyond the prompt
For now, prompts are how we speak to machines. That is the visible change. The subtler one is what we are slowly learning to accept.
Not long ago, an answer came back that was almost perfect. It was fluent, confident, neatly structured. It saved time. It even sounded wise. It was used, lightly edited, and passed along.
Only later did it become clear what it had done.
It had not lied. It had not made anything up. It had simply rounded off the rough edges—the parts where the question was still unresolved, the parts that should have made us uneasy.
By the time this was noticed, the answer had already travelled. It had been forwarded, quoted, absorbed. Correct enough to pass. Shallow enough to not matter.
The mistake was not the machine’s. It had done exactly what it was asked to do.
The mistake came earlier. In the moment the prompt was written. In the decision to accept clarity over accuracy. Fluency over friction.
So the next time, the prompt changed. It became longer. More specific. More demanding. Not because the system needed it—but because we did. Because we had learned, once, what happens when a question is allowed to resolve itself too easily.
People like to talk about a future where this becomes unnecessary. Where systems infer intent perfectly. Where nothing needs to be spelled out. Where the machine simply knows.
But the problem was never misunderstanding.
The problem was our willingness to settle.
We settle for the version of reality that is easiest to digest. But someone still decides when an answer is “good enough.” Someone still chooses when to stop thinking. Someone still lives with the downstream effects of that choice.
Come to think of it, Lewis Carroll was not really talking about roads.
He was talking about what happens when you start moving before you know what you are willing to get wrong.
