How to talk to AI well

Reading time: ~15 min · Prereqs: Lessons 01–04 · 4 interactive demos inside · This is the lesson that changes everything.

Why this matters more than any other lesson

The single most common complaint about AI tools is "I tried it and it gave me useless answers." Almost every time, the problem is not the AI. The problem is the prompt. A well-crafted prompt to a mediocre model often beats a lazy prompt to the best model on the market.

This lesson teaches you prompting — the art of writing messages to AI in a way that gets you useful, specific, ready-to-use output. It's the skill that compounds. Once it clicks, every other AI tool you ever use gets better.

The core idea: AI can't read your mind

Here's the fundamental truth: the AI has no idea who you are, what you're working on, or what "good" looks like to you. It has no memory of you from a previous session (unless a tool specifically adds that). It only knows what's in your current message.

That sounds like a limitation. It's actually an instruction. Give it everything it needs to help you well, right there in your message. The more relevant context and direction you include, the better the result.

A reliable recipe: Role, Context, Task, Format

Researchers and heavy users have converged on a simple structure that works across almost any task. Think of it as the four ingredients of a useful prompt:

• Role — who the AI should act as. "You are an experienced kindergarten teacher." This puts it in the right frame of mind, vocabulary, and perspective.
• Context — the background it needs to help you well. "I'm planning a 30-minute lesson for 5-year-olds about recycling. They have short attention spans." The AI can't ask follow-up questions like a human colleague would. Give it the context upfront.
• Task — exactly what you want it to do. "Give me three hands-on activities." Be specific about scope, quantity, and nature of the output.
• Format — how you want the answer structured. "List each as a numbered step with the materials needed and the time it takes." Without format guidance, the AI defaults to prose paragraphs, which aren't always what you need.

You don't always need all four. A quick casual question can get by with just the Task. But the more the output matters, the more ingredients you should include. The sweet spot for most work tasks is all four.

Start filling in the fields above and your prompt will appear here...

Seeing the difference: vague vs. specific

Reading about the recipe is one thing. Seeing the contrast in action is another. Below are two prompts for the exact same task — one vague, one using the four-part recipe. Toggle between them to feel the gap in what you get back.

Five habits that instantly improve results

The four-part recipe is your starting point. These five habits take you further, and they work alongside the recipe rather than replacing it.

1. Show an example

This is the single most underused trick in prompting. Words like "professional," "friendly," or "concise" mean different things to different people. An example is unambiguous.

Add a line like: "Write it like this: [paste your example here]." The AI will match the pattern, tone, length, and structure of the sample you provide. Use it for emails, social posts, job ads, reports — anything where the style matters.

2. Ask for the format you want

Left to its own devices, the AI defaults to flowing prose. That's often not what you need. Be explicit: "Give the answer as a table with columns for X, Y, Z." Or "100 words max." Or "Format this as a bulleted FAQ." The AI will comply, and the result will be far more useful.

3. Tell it what not to do

Positive instructions alone sometimes aren't enough. Adding a "don't" constraint closes off the paths you don't want:

• "Don't use jargon or technical terms."
• "Don't make anything up — if you're unsure, say so."
• "Don't suggest paid tools."
• "Don't start the email with a pleasantry."

That third one — "don't make anything up" — is worth using any time you ask a factual question. It doesn't eliminate hallucinations entirely, but it dramatically reduces them because the model becomes more likely to flag uncertainty rather than paper over it with confident-sounding nonsense.

4. Iterate, don't restart

New users tend to rewrite their entire prompt from scratch when the first response isn't quite right. That's the slow path. The fast path is to stay in the same conversation and steer:

• "Shorter." — cuts length.
• "More formal." — adjusts tone.
• "Put that in a table instead." — reformats.
• "Now add a section on cost." — extends.

This is exactly like editing a document with a colleague. You don't throw the document away because one paragraph is off — you revise that paragraph. The chat history gives the AI all the context from your earlier turns, so you can build on it rather than repeating yourself.

The animation below shows a real iteration sequence. Hit Replay to watch it again.

5. Let it think for hard problems

For anything that involves multi-step reasoning — planning, comparing options, solving logic puzzles, working through a problem — add one of these phrases to your prompt:

• "Think step by step."
• "Show your reasoning."
• "Walk me through how you arrived at that."

When the model has to write out each step of its reasoning, it makes fewer errors. It's the equivalent of asking someone to show their working in a maths exam. The act of externalising the process keeps the thinking more honest. Don't use it for simple tasks — it makes those unnecessarily verbose. Save it for genuinely complex questions.

Common beginner mistakes (and how to fix them)

You don't have to learn these the hard way. The checklist below covers the five most common traps. Check each one off as you read it — your progress is saved between visits.

When a better prompt isn't the answer

Prompting is powerful, but it's not magic. A few things require more than better wording:

• Genuinely current information — if you need today's news, currency rates, or live data, use a tool with web search. No prompt can give an offline model information it doesn't have.
• Your private documents — the model can't see files it hasn't been given. Paste the relevant section into your message, or use a tool that supports file uploads.
• Precise maths — language models are better at reasoning about maths than at doing arithmetic with guaranteed accuracy. For number-critical work, use a tool with a code interpreter that can execute calculations exactly.

Summary

Prompting is a learnable skill, not a talent. The four-part recipe (Role + Context + Task + Format), the five habits (show an example; ask for the format; say what not to do; iterate; let it think step by step), and awareness of the five common mistakes — that's the whole toolkit. It fits on a single index card. The only thing left to do is practise.

The homework page gives you three scenarios to work through. Each one asks you to write a real prompt, run it, and reflect on what came back. That practice loop is where the skill actually builds.