The Art of Prompting: How You Talk to AI Matters

✦ The Same Question, Two Worlds

Let's begin with a demonstration.

Here are two prompts asking the same thing:

Both are questions about meditation. But they do not produce the same answer.

The first might get you a Wikipedia-style paragraph: broad, safe, unspecific. The kind of answer that technically says something but practically tells you nothing you didn't already know.

The second gives you a focused, expert-voiced, clinically relevant summary. It knows what voice to use, what scope to cover, what length to aim for, and what to leave out.

The knowledge was there in both cases. The prompt decided what got surfaced.

This is not a trick. It is not a workaround. It is simply how the machine works.

✦ Why Your Words Carry More Weight Than You Think

You now know how the model generates output. Token by token. Each choice conditioned on everything before it.

What this means for prompting:

The very first word you type begins shaping the probability space of everything that follows.

"Explain..." primes explanation-shaped continuations.
"List..." primes bullet-formatted responses.
"Imagine you are..." shifts the entire voice and perspective of what comes next.

You are not just asking a question. You are setting the conditions under which the model will operate for the entire response.

Think of it like tuning an instrument before you play. You can play any note after that, but the instrument you pick determines the range of sounds available.

A vague prompt does not produce a bad answer. It produces an average answer. The model, given little to work with, reaches for the most statistically common continuation of those words. Which is: the expected, the generic, the adequate.

Precision is what unlocks specificity.

✦ The Anatomy of a Prompt

Not all prompts are built the same.

The difference between a mediocre prompt and a useful one usually comes down to four things. You don't need all four every time. But knowing all four is what lets you choose.

1. Role

Who is the model speaking as?

The same information sounds completely different from a teacher, a lawyer, a poet, and an engineer. When you give the model a role, you are not just changing the style. You are changing what it decides to emphasise, what it treats as obvious, and what it bothers to explain.

Three roles. The same question about tax filing. Three completely different answers in tone, depth, and priorities.

2. Context

What does the model need to know to help you well?

A doctor who doesn't know your medical history cannot give you useful advice. They can give you general advice, the kind you could get from a pamphlet. The model is no different.

Context tells the model what situation it's operating in. What the user already knows. What matters in this specific case.

The second prompt does not just ask for an email. It gives the model a situation, a relationship, a concern to address, and a tone to strike. The model does not have to guess any of it.

3. Task

What exactly do you want the model to do?

This sounds obvious, but it is the most frequently skipped part of a prompt. People describe a situation and hope the model infers the right action.

Sometimes it does. Often it doesn't.

Be direct about what you want.

Task clarity doesn't just improve the output. It speeds up the whole process. Fewer back-and-forths. Less time spent re-prompting to get what you actually needed.

4. Format

What should the output look like?

Left unspecified, the model picks a format based on what most commonly follows the kind of question you asked. This is often fine. But often you have a preference, and not stating it wastes a step.

Format instructions can be simple:

The model will follow these reliably. They do not constrain what the model knows. They shape how that knowledge comes out.

✦ Not All Four, Every Time

You do not need a 200-word prompt for every question.

"What is the capital of France?" needs no role, no context, no format instruction. The task is clear, the expected output is obvious.

But the moment you're asking for something nuanced, long-form, expert-toned, or domain-specific, the four levers start to matter enormously.

Think of them as available tools, not mandatory steps. A chef doesn't use every spice in every dish. But they know what each spice does.

✦ Before and After

Here is the same request, transformed.

Request: Explain a technical concept

Request: Write something

The gap in output quality between these pairs is not subtle. It is the difference between a response that technically answered you and a response that actually helped you.

✦ Three Common Mistakes

Mistake 1: Asking for everything at once

When a prompt asks for too many things, the model tries to serve all of them. The result is shallow across the board.

Ask for one thing well. Then build on the answer.

Mistake 2: Describing the situation but not the action

"My team is struggling with prioritisation" is context. It is not a task.

The model needs to know what you want done with that context: a framework? a list of questions? a draft agenda? Add the action.

Mistake 3: Assuming the model knows your constraints

The model does not know your word limit, your audience, your deadline, or your tone preference unless you say so. These are invisible to it.

Whatever matters to you about the output, state it explicitly. The model cannot care about what it cannot see.

✦ Takeaway Summary

✦ Try It Yourself

Rewrite these prompts:

Take each weak prompt below and rewrite it using all four levers. Then try both versions in ChatGPT, Claude, or any model you have access to. Note what changes.

For each, ask yourself:

• Who should the model be?
• What context does it need?
• What exactly should it do?
• What should the output look like?

One real test:

Think of something you actually need help with this week. Something real, not a practice prompt. Write the worst version of that prompt: one line, no context. Then write the best version using everything from today. Run both. The gap you see is the skill you're building.

✦ Learn More

• Prompt Engineering Guide (DAIR.AI): comprehensive, community-maintained reference for prompting techniques across all major models
• OpenAI: Prompt Engineering Guide: practical guidance with concrete examples, from the team behind GPT
• Anthropic: Prompt Engineering Overview: Claude-specific guidance, with principles that apply broadly across models