Unit 5Prompt engineering

The big idea

Prompt engineering is getting an AI to put all of its effort into exactly what you want, in the way you want it. The secret is doing your thinking before you hand it over.

Why it matters

An AI's limits usually come from the person using it. Writing a good prompt is the same skill as giving a person a good brief: knowing what you want, what matters, and how you'll judge the result. That skill works everywhere, with or without AI.

My framework: five parts of a good prompt

1. Goal: What do you want?
2. Context: What should it know? Who are you, what's the project, what have you already tried?
3. Constraints: Who is it for? What format? How long? What should it avoid?
4. Examples: What does good look like? Show it one.
5. Quality: How will you check it? Tell the AI how you'll judge the answer.

One-shot prompts

A one-shot prompt puts everything the AI needs, start to finish, into a single message. If you want a website, everything about that website goes in one prompt.

Why not just chat back and forth? Because every extra turn is another chance for a mistake to stack on a mistake (remember the fractions lesson in Unit 4). When you front-load the thinking, the AI spends its effort on the work instead of guessing what you meant.

Do this

1. Pick one of your three questions from Unit 1.
2. Talk it out first. Open an AI with voice mode and say: "I'm going to describe a problem. Ask me questions until you understand it. Don't solve it yet. When you understand it, write me a one-shot prompt I can use to research it." Then just talk. If you can't get your ideas out of your head, saying them out loud helps.
3. Read the prompt it writes. Check it against the five parts. What's missing? Add it.
4. Run the same request in a second AI from a different company. Ask it to write its own one-shot prompt for the same problem.
5. Put the two prompts side by side. When I did this live, one tool gave plain text and the other gave a structured document with a role, research limits, a section on how the idea makes money, and a "graveyard" of products in that space that had already failed. Same request, very different thinking.
6. Combine the best parts into one final prompt. Run it.

Two rules

• More words don't automatically make a better prompt. Clear beats long.
• Judge the AI's feedback too. If it suggests a change, ask yourself whether the change actually helps.

Try it with AI

I want to [goal]. Here's the context: [who you are, what the project is, what you already know]. Constraints: [audience, format, length, what to avoid]. Here's an example of what good looks like: [example]. I'll judge your answer by [how you'll check it]. Before you start, list anything you still need from me.

Questions to ask yourself

• Could a stranger do this task well from my prompt alone?
• Which of the five parts did I skip?
• What did the second AI notice that the first one didn't?

In my words

"You want to spend your time reviewing these prompts, because that's where your best effort is spent nowadays."

"You should never expect more from AI than what you put into it."

Worksheet

Use Worksheet 5: One-shot prompt builder in the PDF kit.