Talking to AI: Your First Prompts

The problem with most prompts

Most disappointing AI output traces back to one thing: the prompt did not contain enough information to answer well. The model is not being lazy. It genuinely could not tell what you wanted, so it produced the blandest thing that fits.

Compare:

The second column is longer because it contains the answers to questions the model would otherwise have to guess.

A structure that works every time

Four parts. You will not always need all four, but when a prompt fails, a missing part is usually why.

1. Role — who is answering. “You are an experienced hiring manager.”
2. Task — the single thing you want. “Review this CV summary.”
3. Context — what you would tell a colleague first. “It is for a junior data analyst role. The candidate is a career changer from teaching.”
4. Format — shape and length. “Give three bullet points of specific feedback, then one rewritten version under 60 words.”

Put together, that is a prompt that can only really be answered one way — which is exactly what you want.

Why “role” helps at all

It is not magic and it is not the model pretending. Naming a role shifts which patterns are most relevant. Text written by hiring managers looks different from text written by marketers, and saying so steers the output towards the right body of material. Modest effect, free to add.

Show, do not describe

This is the highest-leverage habit in the whole lesson.

Adjectives are ambiguous. “Professional but friendly”, “punchy”, “clean” — each means something different to every reader, including the model. An example is unambiguous.

Write product descriptions in this style:

“The Cloud Sneaker — lightweight comfort meets street style. Memory-foam insole. Breathable mesh. Your feet will thank you.”

Now write one for a waterproof hiking boot.

One example fixes tone, length, sentence rhythm and punctuation all at once. Two or three examples are better still.

Iterate; do not argue

A first answer is a draft. The productive move is a small, specific correction:

• “Half the length.”
• “Cut the introduction — start at the second paragraph.”
• “More concrete. Replace every adjective with a fact.”
• “Give me three more like the second one.”

The unproductive move is expressing displeasure. “That’s not what I wanted” contains no information about what you did want, so the next attempt is another guess.

Ask for options

Any single answer is one sample from a large space of possible answers. Asking for several costs almost nothing and gives you something to judge:

“Give me five subject lines. Vary the angle: one curiosity-driven, one direct, one with a number, one asking a question, one blunt.”

You are far better at picking than the model is at guessing your taste. Play to that.

The four common mistakes

• Two tasks in one prompt. “Summarise this and write a reply and translate it” — split it up.
• Assumed context. The model does not know your company, your reader, or last week’s meeting unless you say so.
• Accepting the first draft. It is nearly always the worst one you will see.
• Asking for facts it cannot have. Current prices, live figures, precise citations — those need a source, not a better prompt.

Try this: Take a prompt you have used and rewrite it with all four parts — role, task, context, format. Run both. The gap between them is the whole skill, and it takes about ninety seconds to see.