What a prompt actually is

Ask a librarian “got anything on Rome?” and you’ll get a shrug and a shelf number. Ask “I’m writing 800 words on why the Roman Republic collapsed, for readers who know nothing about it — where should I start?” and you’ll get a conversation. Nothing about the librarian changed. The request did.

The word “prompt” is doing you a disservice

It sounds like a command — something you issue, which the machine obeys. That picture will mislead you for as long as you hold it, because it predicts the wrong failures. If prompts were commands, a clear command would always work, and a refused command would mean the machine was broken.

Go back to E1·L6. A language model continues text. Everything you type — plus the conversation so far, plus any files and any hidden instructions the tool adds — becomes the context it continues from. Your prompt is not an instruction sitting outside the system. It is the raw material inside it.

So the real question is never “what is the magic wording?” It is: what has this model got to work with?

Why phrasing moves the answer

If your prompt is the beginning of a document, then the model is finishing a document that starts that way. “Explain inflation” begins something encyclopedia-shaped, so you get encyclopedia prose. “My rent went up 12% and my salary didn’t — explain what’s happening to me” begins something else entirely, and the continuation follows.

Neither is a trick. You changed what document you were asking to be finished.

This is also why prompts that sound forceful — ALL CAPS, “you MUST”, threats about it being important — do so little. Emphasis is not a lever on the machinery. Specifics are.

What this predicts

Three things follow, and each one saves you time later:

Vague in, generic out. A thin prompt has nothing to make the answer specific to you, so it produces the average of everything similar. That is not laziness on the model’s part; there is genuinely nothing else to go on.
Context beats cleverness. Adding the actual document, the actual constraint, the actual audience does more than any phrasing trick. L3 is entirely about this.
The failure modes from E1 don’t switch off. A beautifully worded prompt still gets confident, wrong specifics. Better prompting narrows the odds; it never converts a plausible-sounding machine into a reliable one.

Try it now (5 minutes)

Pick something you actually need this week. Write the laziest version of the request — five words, no context. Send it. Now write it again with who it’s for, what it’s for, and one real constraint. Send that in a fresh chat.

Read both answers side by side. The gap between them is the entire skill this module teaches, and you just produced it yourself in two minutes.

Check your understanding

1. Mechanically, a prompt is best described as: A command the model is obliged to follow The context the model continues from A password that unlocks better model behaviour

2. Writing 'THIS IS VERY IMPORTANT, YOU MUST BE ACCURATE' typically: Meaningfully increases accuracy Does little — emphasis is not a lever, specifics are Switches the model into a verified mode

3. You ask a thin, five-word question and get a bland, generic answer. Why? The model is being lazy There was nothing in the context to make the answer specific You used the free tier

Recap

A prompt is context, not a command; rewording changes which continuation is likely, which is why specifics beat emphasis; and clearer prompting narrows error without abolishing it. Next: the four parts that turn a vague request into a clear one.