You are not bad at prompting. You are good at talking to humans.

Some years ago the rail industry spent a great deal of money making trains faster. The logic was impeccable. People dislike spending time on trains, therefore reduce the time spent on trains, therefore people will be happier. Hundreds of millions of pounds went into shaving twenty minutes off a journey.

A rather cheaper idea, never seriously considered at the time, would have been to spend a fraction of that on wifi, decent coffee and a comfortable seat — at which point passengers might have asked for the trains to be slower. The engineers were solving the problem as stated. The problem as stated was wrong.

I mention this because almost everyone I meet who is disappointed by AI is making precisely the same category of error. They assume the problem is the model. It is almost never the model.

The politeness trap

Consider what happens when you get into a taxi and say “the station, please.” This is a spectacularly under-specified instruction. You have not said which station, by which route, at what speed, or what you intend to do when you arrive. And yet it works, reliably, thousands of times a day.

It works because the driver has context you did not have to supply. He knows the city. He knows there is only one station worth meaning. He can see you have a suitcase. He has, in short, done ninety per cent of the specification on your behalf, for free, and you have never once thanked him for it.

Human communication is efficient precisely because it is incomplete. Saying more than necessary is not thoroughness; it is rudeness. If you got into that taxi and delivered a four-paragraph brief on your transport requirements, the driver would quite reasonably assume something was wrong with you.

We have spent our entire lives being rewarded for brevity. Then we sit down at a text box and are surprised that the habit betrays us.

So when you type “write me a marketing email” into a language model, you are not being lazy. You are being well-mannered. You are doing the thing that has worked for you in every conversation you have ever had. The trouble is that this particular listener has read most of the internet and knows absolutely nothing about you.

Enormous knowledge, zero context

This is the strange asymmetry nobody warns you about. The model knows more than any colleague you will ever have. It also knows less about your situation than the work experience kid on their first morning.

Ask a new starter for “a marketing email” and they will not produce one. They will ask you six questions first: who is it going to, what are we selling, what happened last time, how formal are we, how long, and what does good look like. Only a fool would answer those questions themselves and call it a briefing document.

The model does not ask. It is congenitally incapable of appearing uncertain. So it guesses — plausibly, fluently, and at length — and hands you the statistical average of every marketing email ever written. Which is, almost by definition, the most forgettable email it is possible to produce.

The opposite of a good idea can also be a good idea. But the average of every idea is never a good idea at all.

What enhancement actually is

Here is where people go wrong in the other direction. Told that longer prompts work better, they write longer prompts — padding, throat-clearing, and the word “please” deployed with increasing desperation. This does nothing. Length is not the active ingredient.

The active ingredient is the context a human would have supplied for you. Role. Audience. Constraints. What to do, and what emphatically not to do. What finished looks like. An enhanced prompt is not a longer request; it is the same request with the taxi driver's local knowledge written back in.

Notice that almost nothing in the second version is creative. It is all context — the things you knew and never said, because in any human exchange you would never have needed to.

The bit that is actually hard

You could of course write that second prompt yourself. You will not. I know you will not, because I do not either, and neither does anyone else. Not because it is difficult but because it is tedious, and tedium is a far more powerful force in human affairs than difficulty.

This is the point most productivity advice misses. People do not fail to do things because they are hard. They fail to do them because there is a small, stupid amount of friction in the way. Remove four seconds of friction and behaviour changes; explain to someone at length why they ought to try harder and nothing happens at all.

• The dishwasher did not save time. It removed the moment of dread before starting.
• Uber's great innovation was not shorter waits. It was the little moving car that told you the wait was ending.
• Nobody reads the manual. Everybody reads the one-line tooltip.

So the useful question is not “should I write better prompts?” — obviously you should, and obviously you won't. The useful question is whether the good version can be made to appear without you having to be virtuous about it.

Why watching it happen matters more than it should

We stream the enhancement as it is written, word by word. From a strictly engineering standpoint this is indefensible: the finished text arrives at the same moment either way. The total wait is identical. We have optimised nothing.

Except that a wait you can watch is roughly half the wait you cannot. This is not a metaphor — it is one of the better-replicated findings in queuing research, and it is why lifts have floor indicators and why the little moving car on the Uber map was worth more than any actual reduction in journey time.

The engineers wanted to make the train faster. The passenger wanted to know when it would arrive. These are not the same problem, and only one of them is cheap to solve.

You also get something genuinely useful, which is the ability to stop. When you can see the thing taking shape, you notice within two seconds that it has misread you — and you can intervene rather than politely reading nine paragraphs of confidently wrong prose out of some misplaced sense of obligation.

The prompt is the asset, not the output

One last inversion, and it is the one most organisations get backwards. Everybody saves the output. Almost nobody saves the prompt.

This is precisely the wrong way round. The output is worth something once. The prompt is worth something every time — it is the reusable part, the compounding part, the part that got better because someone in your team spent an afternoon working out what actually produces good results. And it is currently sitting in their chat history, where it will die.

A shared prompt library is not a filing cabinet. It is closer to a recipe book, in that the value is not the paper but the fact that somebody already made the mistakes. The best thing your most capable colleague can leave behind is not their work. It is the instructions that produced it.

In short

• Your prompts are under-specified because human conversation rewards under-specification. This is a habit, not a defect.
• The model has vast knowledge and no context. Only one of those gaps is yours to close.
• Enhancement supplies context, not length. Padding a bad prompt makes a longer bad prompt.
• You will not do this by hand, because it is boring — and boring beats difficult every time.
• Watching it happen makes the wait feel shorter and lets you kill a bad answer early.
• Save the prompt, not just the output. It is the only part that compounds.

None of this requires you to become a better writer, or to learn a syntax, or to read a forty-page guide on prompt engineering — a genre of document that exists mainly to make its authors feel clever. It requires only that the tedious part happens somewhere other than in your head.