The importance of well-written AI prompts — context is key

The importance of well-written AI prompts — it’s about context, not magic

Everyone uses AI today. But because it’s so new and changes daily, we have to adapt continuously. What does an AI actually do without a properly defined prompt? It starts rambling. It starts making things up. But if you give it the right specifications, it behaves like a true professional.

We’re all beginners — and that’s perfectly fine

I’m not an expert, and I honestly believe no one truly is. AI is learning from us every day — our workflows, our needs. It’s collecting data to improve itself, and we’re learning daily to use it more effectively and to focus it on our day-to-day needs. It’s a mutual learning journey.

I believe a large percentage of businesses already use AI daily: editing emails, automations, or simply finding inspiration when creativity runs dry. One good idea can change the way you see things. We’re all green at this, and we need to adapt daily to keep up with constant innovations.

The AI landscape — who offers what?

The major AI players compete fiercely for customers, offering daily improvements and upgrades across different models. We hear it constantly: “Today Anthropic launched model X,” “OpenAI released its most powerful video model,” “Google launched its new Gemini model,” “Grok wants to implement real-time news.” Every day we face a wide range of choices — but what should you pick? Which one is best?

In reality, I believe it depends on what you’re working on.

Anthropic started with a focus on programming tasks but has gradually improved models for image generation, design, security, automations, and skills. The vision is for users to have their computer connected to Claude and perform all daily tasks via a simple chat — even from their phone. It’s essentially vibe coding in natural language. They offer different models: from basic ones for simple questions and light tasks (around $3 per million tokens with ~200K token context window) to advanced models for programming tasks (~$15 per million tokens).

OpenAI has GPT models with custom GPTs for different tasks, and DALL·E for image creation. Google has Gemini with customised Gems for each task. They all offer roughly the same: coding, image editing, video editing. The same goes for Grok and other players in the market.

The subscription trap — the burger that looked better in the picture

The problem is learning, testing, and exploring each of these tools, seeing how they behave and figuring out which ones suit your tasks. Because in the end you’re paying 3–4 monthly subscriptions to try and test — and the results aren’t always what you expected.

It’s like seeing a photo of a burger at a fast-food place. You say: “Wow, I want that one!” And when you open the box? Surprise — half the size, without the vivid colours from the photo. What a disappointment. Something similar happens with these AI tools.

Benchmarks and statistics — who is actually testing?

You read official documentation about new implementations, comparison tables between models: “This model scored 78% accuracy, the other 82%.” OK, the other one is better. But wait — who creates these statistics? They do themselves, with programs they built themselves, so their tools look better in numbers and stats than the competition. This isn’t a kilo of potatoes you can verify on a scale. But we largely have to trust what they say.

It’s true that there have been clear improvements compared to 2–3 years ago. Things have changed. AI has started reasoning better, understands context, can read images and screenshots, and can help you significantly — if you give it the right context.

Context is everything — the restaurant example

What happens if you talk to AI like a lifelong friend? For example: “Give me the address of the restaurant where we had dinner last week — I can’t remember if it was Wednesday or Thursday.”

Does the AI know where you ate? Does it know if it was Wednesday or Thursday? No. And the answer will match the question: evasive, rambling — it’ll start recommending random restaurants in your area that have nothing to do with you.

But what if you write a prompt about yourself — your habits, your diet, where you eat, your preferences? Then the AI has the necessary information and will answer as precisely as possible.

AI for website building — the brutal reality

The example above was simple. But what happens when someone sees a presentation of a new model with “extraordinary coding abilities” — like Lovable or similar tools that promise websites in minutes without knowing how to code? Influencers doing affiliate marketing for these companies show videos of websites with parallax effects in the hero section, eye-catching colours, spinning and moving objects. “Click the link in the description, create an account, and build your website in minutes.”

You think: “I can do that too! How nice. Let me try.” In theory it sounds great, but practice is the brutal reality.

If you use a Claude model with a basic prompt like “I want a pretty website like I’ve seen online, with moving elements and caramel colours,” and you pick the most expensive model because you think “expensive = quality”... Claude starts asking technical questions: Which technology do you want? React? Next.js? Astro? Questions you can’t answer. And the $15 per million tokens are already gone — with nothing to show for it.

Then you try the tool you saw that “does it all by itself.” You create a Lovable account, pay for the most expensive plan, give it a natural-language prompt, and it creates a simple page with a few basic sections that don’t really describe your business. Maybe visually nice, but it conveys nothing.

Will this page convert customers? No.

The big question: Will this page convert customers? No. Absolutely not. Because:

You haven’t configured a Norwegian domain (.no)
The code is basic — a React app with access to loads of unknown libraries
No typedoc configured with no-nb (language settings)
No pre-defined SEO
No clear heading structure (H1, H2, H3)
No correctly configured robots.txt or sitemap
No way to implement GTM, GA or GSC
No proper copywriting, descriptions or strategic section placement
No contact page linked to your domain — you’d need third-party apps like Formspree, where all customer data passes through them
No proper cookie banner without third-party apps

All this for a simple landing page with practically no purpose, no logic, and no correct structure. You might get a nice design with a good prompt, but you’ll never achieve correct and lasting online visibility.

A simple HTML form converts better than any AI-built page

We guarantee that a simple contact form properly built in HTML and CSS — for example for a plumber in the Bærum area with a phone number — uploaded to index.html on your domain’s server, with a GTM account, GSC, GA, and a Google Maps listing, will convert 1,000 times better than any page built with an AI tool.

AI for application development — a Java perspective

We’re not even talking about those who showcase tools for building web applications on these platforms with promises like “Look, I built an app to manage a driving school in just 5 minutes.” This is something no serious business should even attempt.

We come from Java. We love Java. We’ve learned that any application must have a solid, well-defined structure to function properly. Java is an object-oriented programming language. An app needs a clear, organised, and well-defined structure with a proper separation between backend and frontend.

There are monolithic applications where the Java backend and Thymeleaf frontend live together, and apps where backend is separated from frontend. Every application needs entities, models, repositories, controllers, services, DTOs, database, pom.xml, security, abstract methods, constructors, design patterns, business logic — and much more. We’re not even talking about multi-tenant applications where you need to carefully isolate data to prevent information from different users mixing together.

Doing all this with AI is like asking it to build a car: it builds one, but puts the steering wheel in the boot, the back seats up front, and the gear stick in the rear.

How Kodemagisk uses AI today

Yes, AI can help us — but with the right instructions, the right prompts, and always maintaining control over what it does. You must always verify the changes it makes.

Does Kodemagisk use AI? Yes. We’re at a point where AI makes life easier — but only in certain areas:

Searching documentation quickly and efficiently
Extracting the right information from a 20-page PDF — before, you had to read entire documentation sets to find what you needed
For programming — building complex logic for booking systems or multi-tenancy where AI can guide you well
Replacing hours of searching on StackOverflow and other forums — where you used to browse through threads looking for someone with a similar situation

AI makes many things easier than before. But the key is always the same: give the AI proper context, write precise prompts, and always stay in control of the result.

We hope you enjoyed this article. For questions or enquiries — feel free to get in touch.

📧 Contact us: 📍 Oslo, Norway | Kodemagisk AS