10 Things to Think About When Writing an AI Prompt

There's a whole industry built around "prompt engineering": secret templates, magic phrases, "50 prompts that will change your life." Most of it makes a simple thing sound like a specialist skill. A prompt is just a brief, and if you can explain a job to a person, you can write a decent one.

Here are ten things worth thinking about before you hit enter. Each one has a weak prompt and a better one, and the only difference between them is that one idea. At the end, I'll put a few of them together so you can see what a full prompt looks like.

1. Start with the goal, not the topic

A topic gets you an encyclopaedia entry. A goal gets you something you can use.

2. Say who it's for

The same explanation lands very differently for a ten-year-old, a boss, or your mum. If you don't say, the AI aims at an average reader, which usually means nobody in particular.

3. Define what "good" looks like

"Short" and "professional" mean different things to different people, and the AI will pick its own meaning. Numbers and specifics remove the guesswork.

4. Give it the context it can't guess

It doesn't know your budget, your schedule, or that your kid won't eat mushrooms. Give it the background you'd give a new hire on day one.

5. Show an example

If you've got a style you like, paste it in and say "more like this." One real example beats three sentences describing what you want. It's the difference between explaining how to tie a shoelace and showing someone.

6. Tell it who to be

Asking the same question from a more useful angle changes the answer, even though the question hasn't changed. AI tends to be agreeable by default, so a role is one way to get something more honest than encouragement.

7. Let it ask you questions

Add one line: "Before you answer, ask me anything you need to know." It stops guessing and starts checking, and you'll often realise you hadn't thought of half of what it asks. It's a cheap upgrade for almost any bigger request.

It might come back with something like: "Where are you flying from? How old are the kids? Is the priority relaxing or doing things?" Answer those, and the plan is built around you rather than around "a family."

8. One job at a time

Asking for research, comparison, report and summary in one message usually gets you something shallow at every step. Split it up, and you get to steer in between.

Message 2 only exists because you got to read the answer to message 1. That's the real benefit.

9. Treat the first answer as a draft

The first reply is rarely the finished product, and it doesn't need to be. React to it like you would a colleague's draft. The most common mistake is starting a brand new chat with a slightly reworded prompt, which throws away everything the AI has just learned about what you want.

Two quick nudges usually beat one perfect prompt, and they take less effort than you'd think.

10. Read a bad answer as a clue

When the result is bland, too long, or just off, resist blaming the tool and ask what you left out. The answer is telling you something.

Bland usually means it didn't know who it was for. A wrong tone usually means you didn't show it an example.

It also works the other way. When something sounds confident but feels wrong, ask: "Are you sure about that? What might be wrong with this answer?" It will often give you a more honest second pass. And for anything that really matters, like money, health, legal or anything you're about to put your name to, double-check it yourself.

Putting it together

You won't use all ten every time, and you shouldn't try. Most good prompts use three or four. Here are a few put together, with the numbers they draw on.

An email you need to get right (1, 3, 4, 7)

I need to email my landlord about the heater in the lounge room. It stopped working nine days ago, and I rang last Tuesday and haven't heard back. I want a repair date, not a fight. Keep it polite but firm, under 120 words, and ask for a date by Friday. Before you write it, ask me if there's anything else you need to know.

A weekly plan (1, 3, 4, 8)

Help me plan dinners for the week. There are four of us: two adults and two kids, aged 4 and 7. One won't eat mushrooms and the other won't touch anything saucy. I have about 30 minutes on most nights, plus an air fryer and a slow cooker, and I'm aiming for around $120. Give me five dinners as a list, one line each. I'll tell you which ones I like, and we'll build the shopping list from there.

A reality check on an idea (1, 2, 4, 6)

Act like a sceptical small-business accountant who has watched a lot of ideas fail. I'm thinking about starting a mobile coffee cart at weekend markets for some extra income alongside my day job. Tell me the three most likely ways it loses money and which numbers I should check before I spend anything on a cart. Keep it in plain English, I'm not an accountant.

Something in your own voice (3, 4, 5)

Write a short bio for my dog-walking website, around 60 words. I've been walking dogs around Belconnen for six years and I'm fully insured. Here's one I like for the tone: "I'm Sam. I fix bikes, drink too much coffee, and believe nobody should be embarrassed about not knowing how to change a tyre." Match that relaxed style.

None of these are clever. They're just the details you already had in your head, written down.

How it all fits together

Ten tips is a lot to carry around, and you don't need to. They all come back to one problem: the AI knows nothing about you, your situation, or what you're picturing, and it can't ask unless you let it. Everything above is just a way of closing that gap.

It helps to think of it in three stages.

Before you write, get clear on what's in your head. What am I actually trying to do (1)? Who is it for (2)? What would a good result look like (3)? That takes about thirty seconds, and if you can't answer those, the AI can't either.

While you write, hand over what it can't guess. The background (4), an example (5), a role (6). If you're not sure what it needs, ask it to ask you (7). And keep each message to one job (8), so none of it gets watered down.

After it answers, treat the reply as information, not a verdict. A first draft is something to react to (9), and a disappointing one tells you exactly what you left out (10).

That's the whole loop: say what's in your head, see what comes back, and use the gap to fix the next message. It's the same loop you'd run with a new colleague on their first day. You'd explain the job, they'd come back with something nearly right, and you'd say "close, but..." Nobody would call that a failure to delegate. It's just how working with someone goes.

So the next time an answer is off, skip the "this thing is rubbish" conclusion and ask a better question: what did it not know that I did?