Prompting Basics: How to Talk to AI So It Actually Helps

If you’ve ever typed something into ChatGPT, Claude, or Gemini and gotten a bland, generic, or just-plain-wrong answer, the problem usually isn’t the AI. It’s the prompt — the instructions you gave it. The good news: prompting is a learnable skill, and a handful of simple habits will dramatically improve what you get back.

This is the guide to start with. No jargon you don’t need, and every idea comes with an example you can copy.

What a prompt actually is

A prompt is just the text you send to an AI model. That’s it. But here’s the mental model that changes everything:

The AI is an incredibly capable, well-read assistant who has no idea who you are, what you’re working on, or what “good” looks like to you — until you tell it.

It won’t ask clarifying questions unless prompted to. It won’t read your mind. It fills in the gaps with the most average answer that fits your words. So a vague prompt gets a vague, average answer. A specific prompt gets a specific, useful one.

Compare:

Write a product description.

versus

Write a 60-word product description for a stainless-steel insulated water
bottle aimed at hikers. Emphasize that it keeps drinks cold for 24 hours.
Tone: rugged but friendly. End with a short call to action.

Same AI. Wildly different results. The second one removed the guesswork.

The anatomy of a good prompt

Most strong prompts have some combination of five parts. You don’t always need all five — but knowing them gives you a checklist when an answer disappoints.

1. Role — who the AI should be

Telling the model to adopt a role focuses its “voice” and priorities.

You are an experienced pediatric nurse explaining things to worried parents.

2. Task — what you want it to do

Be a verb-first and concrete: summarize, rewrite, compare, brainstorm, critique, translate, draft.

Summarize the text below into 5 bullet points a busy manager could skim.

3. Context — the background it needs

This is the highest-leverage part and the one people skip most. Give it the raw material: the document, the audience, the constraints, the goal.

Context: This email is going to a client who is upset about a late delivery.
We were at fault. We want to keep the relationship. Our policy allows a 15% refund.

4. Format — how the answer should look

Respond as a markdown table with columns: Option, Pros, Cons, Cost.

5. Examples — show, don’t just tell

One or two examples of the output you want is often worth more than a paragraph of description. (This is called few-shot prompting — more on it below.)

Put together, a full-anatomy prompt looks like this:

You are a senior copywriter. (role)
Rewrite the paragraph below to be punchier and 40% shorter. (task)
It's the intro to a blog post for first-time home buyers who feel overwhelmed. (context)
Keep it warm and reassuring; no jargon. (context)
Return only the rewritten paragraph. (format)

Paragraph: [paste here]

Core patterns worth knowing

Few-shot prompting (give examples)

Show the model the pattern you want it to follow.

Turn these features into benefits. Follow the examples.

Feature: 10-hour battery -> Benefit: Work all day without hunting for an outlet.
Feature: 256-bit encryption -> Benefit: Your data stays yours, even on public Wi-Fi.
Feature: 2-minute setup ->

The model will complete the last line in the same style. Few-shot is the single fastest way to control tone and structure.

Chain-of-thought (ask it to reason step by step)

For anything involving logic, math, or multi-step reasoning, ask the model to think through it before answering.

A store sells notebooks at 3 for $12. If I buy 7, what do I pay?
Think step by step, then give the final number on its own line.

Making the reasoning explicit catches errors that a snap answer would hide. (Note: the newest “reasoning” models do a lot of this internally — but the instruction still helps on tricky problems.)

Constraints (fence it in)

Tell the model what not to do, and set hard limits.

Explain how vaccines work in under 100 words, for a 10-year-old.
Do not use the words "antigen" or "pathogen."

Constraints like word counts, reading levels, and banned words are some of the most reliable levers you have.

Iteration (treat it like a conversation)

You almost never get the perfect answer on the first try — and you’re not supposed to. The real skill is steering:

• “Shorter, and more casual.”
• “Good, but keep the second point and drop the rest.”
• “Now give me three more options that take a completely different angle.”

Each reply is context the model keeps building on. Prompting is a dialogue, not a slot machine.

Do’s and don’ts

Do:

• Give context generously. Paste the document, the data, the earlier draft.
• Say who it’s for. “Explain to a beginner” vs. “for a technical audience” changes everything.
• Ask for a specific format. Bullets, table, JSON, email — name it.
• Ask for options. “Give me 3 versions” beats one take you have to accept or reject.
• Tell it to ask you questions if something’s unclear: “If you need more info to do this well, ask me first.”

Don’t:

• Don’t be vague and then be surprised. “Make it better” — better how?
• Don’t cram five unrelated tasks into one prompt. Break them up.
• Don’t assume it knows private facts — your company’s numbers, today’s news, your files — unless you provide them.
• Don’t blindly trust confident answers. Models can be fluently wrong (“hallucinate”). Verify anything that matters.

A reusable starter template

Keep this around. Fill in the blanks and you’ve got a strong prompt every time:

You are [role].
Your task is to [task].
Context: [everything relevant — audience, goal, constraints, source material].
Format the answer as [format].
If anything is unclear, ask me before you start.

Where to go next

You now know more about prompting than most people who use AI daily. The rest is reps.

• Grab ready-made prompts from the prompt library and study why they’re structured the way they are.
• Try templates — the same idea with fill-in-the-blank {{variables}}.
• When you’re ready to build things with AI instead of just chatting, read Vibecoding.

The best way to get good is to notice when an answer disappoints, and ask yourself which part of the anatomy was missing. Usually it’s context. Add it, and try again.