A prompt is not just text — it's an instruction

In the world of artificial intelligence, we often focus on choosing the right language model — GPT-4, Claude, Gemini. But scientific research shows that prompt quality can be just as important, and sometimes even more important than the model itself. How we formulate a command directly translates into response quality, precision, and usefulness of results.

Scientific research: half the success is the prompt

A study conducted by a research team associated with MIT Sloan revealed something surprising: using a newer version of an AI model accounts for only about half of the observed performance improvement. The other half of success resulted from users skillfully creating and adjusting prompts.

In the experiment, participants were tasked with recreating a reference image using commands issued to artificial intelligence. Users employing the more advanced DALL-E 3 model crafted prompts that were 24% longer and more descriptive than those using DALL-E 2, which directly translated into better results.

This shows that having a better model isn't enough — you need to know how to use it.

Chain-of-Thought: how prompt structure affects diversity

A study conducted by Lennart Meincke, Ethan R. Mollick, and Christian Terwiesch, published on arXiv, focused on increasing the diversity of ideas generated by AI. Using the GPT-4 model, researchers discovered that employing the "Chain-of-Thought" (CoT) technique in prompts leads to greater diversity and uniqueness in generated ideas.

Chain-of-Thought encourages the model to present its reasoning process step by step, resulting in more thoughtful and varied responses. This isn't just a matter of aesthetics — it's a real difference in quality and usefulness of results.

Systematic review: the importance of prompt design

A systematic review of prompting methods in natural language processing, published on arXiv, emphasizes the importance of proper prompt design. The authors point out that precise command formulation allows for better adaptation of language models to new tasks, even with limited training data.

This is particularly important in a practical context: we often don't have the ability to retrain a model, but we can significantly improve its performance through better prompts.

Practical implications

1. Investment in prompt engineering pays off

If you're spending money on a better AI model but not investing time in learning to create good prompts, you might be losing up to half of your tool's potential.

2. Length and detail matter

Research shows that longer, more descriptive prompts often yield better results. This isn't about "more is better" — it's about giving the model enough context and instructions.

3. Prompt structure affects quality

Techniques such as Chain-of-Thought, few-shot learning, or Generated Knowledge can significantly improve results. It's worth knowing and consciously applying them. Below you'll find brief explanations of these techniques.

4. Prompt engineering is a skill

This isn't something you can learn in five minutes. It's a skill that requires practice, experimentation, and understanding how language models process information.

Prompting techniques — brief explanations

Here are the most important techniques worth knowing:

Chain-of-Thought (CoT)

Instead of asking the model for a direct answer, we ask it to show its reasoning step by step.

Example:

• ❌ Poor prompt: "What is 15 × 23?"
• ✅ Chain-of-Thought: "Calculate 15 × 23, showing each step of the calculation."

The model will then show: "15 × 20 = 300, 15 × 3 = 45, so 300 + 45 = 345", which leads to more precise results.

Few-shot Learning

We show the model several examples of what we expect before asking it to perform the task.

Example:

Translate to formal language:
Informal: "Hey, what's up?"
Formal: "Good day, how are you?"

Informal: "Hi, what's going on?"
Formal: "Hello, how are you feeling?"

Informal: "Yo, what's up with you?"
Formal: "Good day, how are you?"

The model learns the pattern from examples and applies it to the new task.

Generated Knowledge

First, we ask the model to generate knowledge/facts on a given topic, then we use that knowledge to provide an answer.

Example:

Step 1: "List 5 most important facts about climate change."
Step 2: "Based on the above facts, explain why reducing CO2 emissions is important."

This leads to more well-argued and precise answers.

Role-playing

We assign the model a specific role or perspective from which it should respond.

Example:

• ❌ "Explain how blockchain works."
• ✅ "You are an experienced programmer explaining blockchain to a beginner. Use simple analogies and examples."

The model adjusts its style and level of detail to the assigned role.

What this means for you

If you work with AI — whether as a programmer, analyst, or entrepreneur — it's worth:

• Investing time in learning prompt engineering — it can be just as important as choosing a model
• Experimenting with different techniques — Chain-of-Thought, few-shot examples, role-playing
• Testing and iterating — one prompt is rarely perfect the first time
• Documenting what works — build a library of effective prompts for your tasks

Summary

Scientific research clearly shows: prompt quality matters enormously. It can be just as important as choosing a language model, and sometimes even more important. This isn't just a matter of "art" — it's something that can be measured, tested, and improved.

It's worth treating prompt engineering as a serious skill that can significantly impact the efficiency of working with AI. Because the latest model with a poor prompt can yield worse results than an older model with a well-designed prompt.

If you're wondering whether AI is even needed in your project, also read: Is AI Always Needed?

Need help learning prompt engineering?

If you want to develop your prompt engineering skills and learn to create effective prompts, I'm here to help. I can assist you with:

• Understanding how different prompting techniques work
• Learning to create better prompts for your specific tasks
• Optimizing existing prompts to achieve better results
• Building a library of effective prompts tailored to your needs

Feel free to reach out — together we can significantly improve the efficiency of your work with AI.

Sources

1. Jahani, E., Manning, B. S., Zhang, J., et al. (2024). As Generative Models Improve, People Adapt Their Prompts. arXiv preprint arXiv:2407.14333
2. Meincke, L., Mollick, E. R., & Terwiesch, C. (2024). Increasing Idea Diversity in AI-Assisted Ideation. arXiv preprint arXiv:2402.01727
3. Liu, P., et al. (2021). Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing. arXiv preprint arXiv:2107.13586