Prompting 101: How to Talk to AI

The reason AI hands you a bland, useless answer usually isn't the model — it's the prompt. Here's the mechanism in plain English, and one real prompt taken from bad to good, one change at a time, so you can watch each fix work.

You typed a question into ChatGPT or Claude, hit enter, and got back something that sounded fine and helped with nothing. Confident, well-formatted, generic. You closed the tab thinking the tool was overhyped.

Here's the moment it usually goes wrong. Someone asks the AI, "Give me some tips for my presentation," and gets ten tips so bland they'd apply to any presentation ever given: "Know your audience." "Practice beforehand." "Use visuals." All true, all useless, because the AI had no idea what the presentation was, who it was for, or what the person was actually stuck on. The answer was generic because the question was.

That gap is the whole subject of this piece. Prompting isn't a secret art with magic words — it's clear communication with a system that takes you literally. And once you see why a vague prompt produces a vague answer, fixing it stops being trial and error.

Why vague questions get vague answers

To fix your prompts, you need one idea about how the tool works under the hood — nothing technical.

A large language model answers by predicting, over and over, the most likely next chunk of text given everything you've written so far. That's genuinely most of what it does. (If you want the fuller picture, what large language models actually are walks through it.) The practical consequence is what matters here: your prompt is the entire set of clues the model uses to decide what comes next. It can't see your screen, your job, or what's in your head. It only has your words.

So picture what "predict the most likely continuation" does with a thin prompt. Ask "Tell me about marketing," and there are a million reasonable continuations — a history, a definition, a listicle, a strategy. With nothing to narrow it down, the model drifts toward the most average answer, the one that fits the widest range of things you might have meant. Average is exactly what "generic" feels like.

Now add clues. Every specific detail you put in the prompt quietly deletes a huge pile of possible answers. Name the audience, and every answer aimed at the wrong audience becomes unlikely. Name the format, and every wrongly-shaped answer drops away. You're not casting a spell. You're steering the prediction by removing the room to guess.

That single mechanism drives everything below. Three kinds of clue do most of the steering:

1. The task — what you actually want done, stated as an action, not a topic.

2. The context — what the model can't know unless you tell it: audience, situation, constraints.

3. The format — the shape of the answer: an email, three bullets, a table, a paragraph.

Leave any one out and the model fills the gap with an average guess. Everything from here is just those three, made concrete.

One prompt, fixed one change at a time

Rules slide off. Watching a single prompt improve sticks. So here's one real, ordinary task — the kind you'd actually bring to an AI — taken from a bad prompt to a good one, one change per step. (Developed example.)

The task: you need to email a coworker to push back on a deadline. You're slammed, the date they want isn't realistic, and you don't want to sound like you're dropping the ball.

Version 1 — the vague ask

Write an email about a deadline.

The model has almost nothing. Which deadline? Whose? Pushing it out, confirming it, chasing someone else's? "Predict the most likely email about a deadline" lands on the blandest possible template:

Subject: Regarding Our Upcoming DeadlineDear Team, I am writing to touch base regarding the upcoming deadline...

It's addressed to no one about nothing. Not because the model is dumb — because you gave it nothing to be specific about.

Version 2 — add the task and the context

Now say what you actually want and hand over what the model can't know:

Write an email to my coworker Priya asking to move a deadline.
She asked for the report by Thursday. I can realistically get it
to her by Monday because I'm covering for a teammate who's out sick.
I want to sound reliable, not like I'm making excuses.

This is the big jump, and it's worth seeing why it's big. "Asking to move a deadline" fixes the task. Priya, Thursday, Monday, covering a sick teammate — that's context the model had no way to invent, and each detail kills off a category of wrong answers (every email to the wrong person, about the wrong dates, with the wrong reason). "Reliable, not excuses" steers the tone. The reply is now a real, sendable email that names the dates and gives the reason.

Version 3 — add the format

The email's good but longer than you'd send to a busy colleague. So specify the shape:

[same as above]
Keep it under 80 words, warm but direct, and propose Monday as a
firm new date rather than asking if it's okay.

"Under 80 words" and "propose a firm date rather than ask" change the structure, not just the wording. You've now steered length, tone, and even the negotiating stance. This is the step people skip most — they specify what to write but never the shape, and then wonder why the answer is three paragraphs when they wanted three lines. The model isn't guessing your format anymore because you stopped making it guess.

Version 4 — show an example (when words aren't enough)

Sometimes you want a specific style that's genuinely hard to describe — so show it instead of describing it. Paste a message you've sent before that sounds like you:

[same as above]
Match the tone of this email I sent last month:
"Hey Sam — quick one. I hit a snag with the vendor export and
want to get it right, so I'd rather send it Wednesday than rush
something broken today. Flagging early so nothing's a surprise.
Thanks for the patience!"

This is "few-shot prompting," and it's the whole trick: one concrete example beats a paragraph of adjectives. You could write "warm, casual, proactive, slightly informal" and still not land it — but one real message shows the model exactly the register you mean, and it matches it. Demonstration beats description because the example is the pattern, with no translation lost in between.

Line up the four versions and the point is hard to miss: nothing about the model changed between "useless" and "exactly right." Only the prompt did.

Version	What you added	Why the answer improved
1. "Write an email about a deadline"	Nothing	Model guesses; lands on the blandest average
2. + task and context	Names, dates, reason, goal	Kills every answer about the wrong thing
3. + format	Length, tone, stance	Fixes the shape, not just the words
4. + an example	A message in your voice	Shows a style words can't fully describe

When a better prompt won't save you

One honest limit, so you don't waste an afternoon rewording a prompt that was never going to work.

A sharper prompt removes the model's room to guess. It cannot give the model knowledge or abilities it doesn't have. If you ask about something that happened after its training cut-off, or about your company's internal systems it has never seen, no amount of rewording conjures the facts — a confident, wrong answer is still wrong. And some tasks sit outside what these tools do reliably at all: exact arithmetic across long chains, guaranteed-accurate lookups, anything where being subtly-wrong-but-fluent is dangerous. That boundary is real and worth knowing up front — what LLMs can't do, and why that matters maps it, and the practical guide to AI sets the wider context.

The skill isn't only writing good prompts. It's telling a prompt problem (fixable by being more specific) apart from a tool problem (fixable only by using a different tool). Reword the first. Stop rewording the second.

The move to take with you

Here's the one habit that turns inconsistent results into reliable ones: when the answer is bad, assume the prompt is the problem first — then read your prompt back as if you were a stranger who knows nothing about your situation. Can that stranger tell what the task is? Do they have the context you have in your head but never typed? Do they know what shape the answer should take?

Nine times out of ten you'll find the gap yourself — a missing detail, an unstated audience, a format you never named — and adding it is the whole fix. That's the good news buried in "AI gives me useless answers": the lever is in your hands, not the model's. Name the task, hand over the context, specify the format, and show an example when words fall short. Do that and you're not whispering incantations at a black box. You're just briefing it well.

Related reading

• What Are Large Language Models? — the "predict the next chunk" mechanism this piece leans on.

• What LLMs Can't Do (and Why That Matters) — knowing when a better prompt won't help.

• What Is Artificial Intelligence? A Practical Guide — the wider map this sits inside.

Keep learning. This article is part of the Start Here path in the ShiftQuality Learning Center. New to quality and delivery? This is the place to begin.

Frequently Asked Questions

How do I write a good prompt for ChatGPT or Claude?

Answer three questions in your prompt: what you want (the specific task, not a vague topic), what context the model needs (audience, domain, constraints), and what format the output should take (list, paragraph, code, email). When you get a generic answer, it's almost always because one of those three is missing — add it and the answer sharpens.

Why does the AI give me such generic answers?

Because a vague prompt leaves the model free to guess, and it guesses toward the most average, most common answer in its training. 'Tell me about marketing' could be answered a thousand ways, so you get the blandest one. The fix is to remove the room to guess: name the task, the context, and the format you actually want.

What is few-shot prompting?

Showing the model one or two examples of the input-and-output pattern you want, instead of describing it in words. Demonstration beats description for anything format-sensitive — paste an example of the style or shape you want, then ask for the same on your real input.

Do I need to learn special prompt-engineering tricks or magic words?

No. Prompting is clear communication, not incantation. If you can brief a capable stranger who knows nothing about your situation, you can prompt well. The 'tricks' (roles, examples, step-by-step) are just structured ways of being more specific about the task, the context, and the format.