Prompt engineering is a skill that emerged along with designing AI features. In small teams, like Copyfolio’s,  system prompts are the result of strong designer-developer collaboration. 

However, in the future, it’ll be a skill every designer will need to take on. Here’s a simple, no-jargon guide to help you write a great prompt that actually works.

1. Learn descriptive writing

The foundation of a great prompt is how well you can describe the task the AI needs to do. It sounds simple, but you’d be surprised how a single missed step or poorly chosen word can make or break the prompt.

You can easily practice this by taking a simple example or action that you’ve done a thousand times. For example, if you were to describe how to make a PB&J sandwich, you’d probably say something like this:

1. Lay the bread flat.
2. Take the peanut butter and spread it onto it.
3. Take another piece of bread.
4. Take the jelly and spread it onto it.
5. Lay the first piece on top of the second one.

The problem is, this isn’t descriptive enough. If you really think about it, when you’re making a PB&J, you don’t just “lay the bread flat.” You take a piece of bread out of the bag, then lay that piece of bread flat, get a knife, use that to spread the peanut butter, and so on.

Try to think about AI actions the same way when writing your prompt. What’s a step zero you’re forgetting? What’s something the AI needs to complete that action?

Remember that AI does exactly what we say, not necessarily what we mean—especially in an environment where it may lack context.

2. Add all the necessary context

When building a system prompt, there’s a framework you can follow that works in most cases:

1. Describe the main task – “create a weekly meal plan”
2. Add a persona – “act as a nutritionist”
3. Add the output format – “organize the data into a table”
4. Add context – “we're a family of four, vegetarians, and we have limited time to cook”

Balancing context is the most important part of your prompt. You’ll need to iterate and, through that, find the fine line where the context is just enough for the AI to achieve its main goal—but you don’t want to confuse it.

A great, real-life example of this is when we built Brandi, the branding coach that lives within Copyfolio, a website builder platform by UX studio. Brandi has a chat interface and helps users define the foundations of their brand, like their USP, brand personality, and tone of voice.

We didn’t want it to wrap up the conversation after the foundations were defined, so we added one additional task: “give them actionable tips on how to update their website,” and left it at that. Still, it gave great advice, like “incorporate your USP into your headline” or “update your About Me section with your strengths.” This might not have been the case if we had confused it with tips on website building.

So when you’re adding context, remember to balance what’s needed and what’s not. Think clarity over quantity—more context helps, but only if it’s relevant.

3. Structure it as you would a notebook

One thing that surprised me when I started learning about prompt engineering is how lengthy system prompts can be, especially when it comes to more complex AI features or actions.

Long prompts themselves aren’t an issue, but you have to learn how to structure them properly. The good news is, it’s quite simple: use markdown, and structure everything exactly as you would in your own Notion or Google Doc.

• Use # hash symbols, ## for different headings.
• Use numbered lists when the order is important.
• Use bulleted lists when giving examples.
• Use bold text for emphasis.
• Use ALL CAPS for restrictions.

Just like humans, AI will get lost in a wall of text. A long and unstructured prompt is going to be just as confusing as a short prompt that’s missing context. However, longer prompts don’t necessarily work better than shorter ones. You can leave a few tasks up to the “imagination” of the AI—but you can only learn this through experimenting.

4. Break down tasks into subtasks

There are things you never want to leave up to the AI’s “imagination,” and that’s the exact breakdown of tasks it needs to do to achieve its main goal. Don’t just tell the AI what to do—using your brand-new descriptive writing skills—but also teach it how to “think” to get there. This means guiding its process step by step, just like you would if you were delegating the task to a human.

Using the same example of creating a weekly meal plan, this is how it could look:

How to think:

1. Start by listing all vegetarian ingredients that are easy to find and quick to prepare.
2. Group them into meals that balance proteins, carbs, and vegetables.
3. Make sure no two days feel repetitive.
4. Consider time constraints: include faster recipes on weekdays and slightly more complex ones on weekends.
5. Review the plan for nutritional balance and modify it if needed.
6. Finally, add a grocery list.

See how this mirrors what you would do yourself: breaking down complex tasks into smaller subtasks. This helps AI reason more like a human expert, resulting in more reflective outputs.

5. Add your own examples

If I were to give anyone one single piece of advice about prompt engineering, it would be to add your own examples and references. This is something we learned early on—AI handles and learns from these very well. Usually a lot better than if we were to simply describe the examples to it.

What does this mean? To use the same meal-planning example, you could:

• Attach a recipe or two that you really like.
• List a few ingredients that you always have at home.
• Attach a meal-planning template you’d like it to use.

The more examples and references, the better.

6. Set clear boundaries

You might notice that AI, depending on the model you’re using, will sometimes do things you didn’t ask it to do—like give extra examples or ask follow-up questions to keep the conversation going, even if you didn’t include that in the prompt.

In these cases, you can add a separate section to the prompt to specify what the AI shouldn’t do. That’s exactly what we did with Brandi’s prompt when we noticed it added example answers right next to every question it asked the user. It made the interaction feel more like a quiz or test, rather than a natural conversation.

Keep in mind that this won’t be necessary for every prompt—you’ll only discover when you need it through testing and iteration.

7. Test, refine & iterate

Just like with any design process, iteration is going to be a huge part of prompt engineering as well. You’ll probably never get it right on the first try—and that’s okay. No one does.

However, you need to be mindful of the process you’ll go through when writing the prompt. We, as a small team at Copyfolio, quickly needed to adapt our processes when we started building AI features. The regular double diamond design process didn’t work anymore, as we quickly realized we needed to work on the first few steps simultaneously. This means the first design sketches will affect your initial prompt, and vice versa.

Once you’ve sketched the first screens and have an idea of what you’d like to do, you can easily test your prompt using ChatGPT or your AI assistant of choice—but try to stick to the model you’ll actually be using. Simply send your system prompt to a new chat and see how the AI reacts. This is a great way to iterate on its tone, for example, if needed.

Once you have the first version of the design and the prompt, only then can you move on to develop a proof of concept and test it with real users.

Closing thoughts

Prompt engineering might sound technical, but at its core, it’s just another design skill. You’re shaping inputs to shape outputs—something designers do every day.

So treat prompts like you would design tasks: start simple, structure them well, give the right context, set clear boundaries, and always iterate. With that mindset, you’ll design AI features that make sense, feel right, and actually do their job.