How to Write Better AI Prompts: 9 Techniques That Actually Work

I used to think the AI was the problem. I'd ask for something, get back a bland paragraph that technically answered the question and helped me with nothing, and quietly decide the tool was overrated. Then I started reading my own prompts back. "Write me something about marketing." Of course that's what I got. I gave it nothing and it gave it right back.

So here's the thing nobody really needs to hear framed as a course: you don't need "prompt engineering." You need a handful of habits. Nine of them, in my experience. Each one below has a copyable example and a before/after, because I'd rather you see the difference than take my word for it.

1. Give it a role

Tell the AI who it's supposed to be and it shifts its vocabulary, its assumptions, and how much it explains before it writes a single word. This is the cheapest win on the list.

Before:

Explain compound interest.

After:

You are a high school economics teacher explaining compound interest to a 15-year-old who finds math boring. Use one everyday example and avoid jargon.

The first answer reads like the textbook nobody finished. The second picks a concrete example and stays at the right level. If you only change one thing about how you prompt, change this.

2. Add context, not just instructions

The AI doesn't know your situation. It can't. So the gap between a generic answer and one that actually fits your problem is almost always context you didn't bother to include.

Before:

Write a reply to this customer.

After:

Write a reply to this customer. Background: they bought our software two days ago, it crashed, and they're asking for a refund. We want to keep them as a customer and can offer a free month. Tone: apologetic but not groveling. Keep it under 120 words.

Look at how much of that "After" is just facts. What happened, what you want, what you can offer. The actual instruction — "write a reply" — is the smallest part of it.

3. Be specific about the output you want

Left to its own devices, the model hands you a paragraph of medium length, every time. If you want anything else, you have to say so. Format, length, structure — spell it out.

Before:

Give me ideas for a podcast.

After:

Give me 10 podcast episode ideas about personal finance for people in their 20s. Format as a table with three columns: episode title, one-sentence hook, and the main takeaway. Keep titles under 8 words.

The table constraint by itself turns a wall of text into something you can actually skim.

4. Show an example (few-shot prompting)

When I want a particular style, describing it in words is a coin flip. Showing one example and saying "do it like this" works almost every time. The fancy name is few-shot prompting. The idea is just: demonstrate, don't explain.

Example:

Rewrite these product names in a playful, punny style. Here is the style I want:

Input: "Lemon Soap" -> Output: "Easy Squeezy Lemon Soap"

Now do the same for: "Mint Toothpaste", "Coffee Mug", "Wool Socks".

One worked example teaches the pattern faster than three sentences of me trying to define "playful."

5. Ask for step-by-step reasoning on hard problems

Anything with logic, math, or a few decisions stacked on top of each other — ask the model to work through it step by step. It slows down in a way that cuts down on the careless mistakes.

Before:

A shirt is 30% off and then I use a $10 coupon. Original price $80. Final price?

After:

A shirt is 30% off and then I use a $10 coupon. Original price $80. Work through it step by step, showing each calculation, then give the final price on its own line.

The bonus is the trail. When the final number looks off — and it sometimes will — you can scroll back and see exactly where things went sideways instead of just distrusting the whole answer.

6. Constrain the format hard when you'll reuse the output

If the answer is headed somewhere specific — a spreadsheet, an email, a slide — tell the model where it's going.

Example:

Summarize this article in exactly 5 bullet points. Each bullet must be one sentence, under 15 words, and start with a verb. No intro, no conclusion, just the bullets.

Hard limits ("exactly 5", "under 15 words", "no intro") stop it from padding. That last bit is the one I lean on most. Models love to wrap everything in a friendly little intro and a "hope this helps" outro, and "no intro, no conclusion" is the only thing that reliably shuts that off.

7. Ask the AI to ask you questions first

This is the technique almost nobody uses and the one I'd fight to keep. When you're not even sure what you want, flip it around. Make the model interview you before it answers anything.

Example:

I want to plan a 5-day trip to Japan. Before you suggest an itinerary, ask me up to 5 questions about my budget, interests, travel style, and constraints. Wait for my answers before planning.

What you get back is an itinerary built around you instead of the same tourist loop everyone gets. It works for resumes, business plans, study schedules — anything where the right answer depends on details that only live in your head. If you want to go deeper on this style, the techniques in our brainstorming prompts guide pair nicely with it.

8. Give it a persona and a point of view

A role (technique 1) tells the AI who it is. A persona tells it how to think — what to push on, what to care about, what to ignore.

Example:

Act as a tough but fair startup investor reviewing my idea. You care about realistic revenue and you're skeptical of hype. Here is my pitch: [paste pitch]. List your three biggest concerns and what would change your mind.

"Tough but fair," "skeptical of hype" — that's you switching off the model's default agreeableness. And the default agreeableness is exactly where the genuinely useful feedback goes to hide. Most AI happily tells you your idea is great. You have to ask it not to.

9. Chain prompts instead of cramming everything into one

Complicated tasks go better as a sequence of small prompts than one enormous one. Each step builds on the last, and — this is the real benefit — you can fix course between steps instead of discovering at the end that it misread you on line one.

Here's a chain I'd actually use to write a blog post:

1. "Give me 10 angle ideas for a post about home composting for apartment dwellers."
2. (Pick one.) "Outline that angle as 5 sections with one-line descriptions."
3. (Edit the outline.) "Write section 2 in a friendly, practical voice, about 150 words."
4. "That paragraph is too salesy. Rewrite it more plainly, no exclamation marks."

You stay in the driver's seat the whole way, and the result beats whatever "write me a blog post about composting" would've spat out in one go. It's not close.

Putting it together

You don't need all nine in a single prompt. Most of my everyday prompts use three or four — a role, some context, an output format, a constraint. Stacked, it looks like this:

You are an experienced copywriter (role). I run a small bakery and want to announce a new sourdough loaf to my email list (context). Write a 90-word announcement (constraint) with a warm, local-shop tone. End with one clear call to action and don't use the word "delicious" (constraints).

And when the first answer isn't quite right? Don't start over. Iterate. Tell it what to fix: "shorter," "warmer," "drop the second paragraph," "make the call to action a question." Needing a second pass doesn't mean you wrote a bad prompt. That's just how this works.

One honest caveat, because I'd feel bad leaving it out: better prompts make AI more useful, not more correct. A beautifully phrased prompt will still hand you a confident, wrong answer — especially on facts, dates, and numbers. Treat what comes back as a strong first draft, not a verified one, and check anything that actually matters.

Every technique here you can try for free, no signup, on Smillee AI. Paste a prompt, see what comes back, refine it. Honestly the fastest way to get better is to run twenty prompts and start noticing which phrasings keep getting you what you want. You learn it by doing it, not by reading about it — including this.

— Maya