How do you ensure good output when using generative AI

#4.1Prompt engineering: how do you steer the AI?

Prompt engineering is a skill that, since the rise of language models, has grown into both a field of research and a set of practical skills. It involves effectively formulating an input or instruction text (the prompt) for an AI model so that the generated output is relevant, useful, and aligned with your goal. A prompt is really nothing more than an assignment or question, but one that you must be able to ask precisely and purposefully. A good prompt therefore includes clarity, context, and structure. Just like asking good questions, which is also an important academic skill. Research by Zamfirescu-Pereira et al. (2023) shows that many users struggle with creating good prompts, especially when they are inexperienced or receive little feedback. That is why you can read more about prompt engineering below.

There are several basic principles that help you write effective prompts. In doing so, it is important to:

1. Determine your goal: for example, do you want an explanation, analysis, summary, creation, feedback, or coaching? If, for example, you use AI to create a summary of a lecture video, you choose a different approach than when you want feedback on a first version of your thesis introduction. When preparing for an exam, you can have practice questions generated or ask for explanations in very simple language.
2. Give the AI a role, for example: teacher, researcher, critic, journalist, fellow student, or end user of a product.
3. Provide the relevant context, such as: target audience, background information, and sources.
4. Determine the type of output: language, number of words, short, in bullet points, in academic style, with formatting (MarkDown), image, diagram, table.

Prompts with clear steps or lines of reasoning (such as 'first explain this, then give three examples') lead to better and more consistently useful results (Liu, 2021; Wei, 2022). You can also ask the AI, as an intermediate step, what information is still missing to carry out the task as well as possible. Or suggest that the AI ask you all the questions it needs in order to carry out your task as well as possible.

Prompt engineering is therefore more than just asking something. It is a step-by-step process in which you keep adjusting your prompt based on the generated output (Brown, 2020; Kojima, 2022). In fact, it actually resembles learning to argue: through language, you try to give direction to a system that has no intentions or goals of its own, but does statistically learn which word order is logical.

Good prompts are therefore clear, specific, and aligned with the goal. So do not ask only: "What is AI?", but rather: "Give a comparison of three types of AI models aimed at educational applications in higher education, including advantages and disadvantages." Or: "Summarize this article in 200 words for a second-year psychology student, with emphasis on the study’s conclusion."

Prompt engineering is the skill with which you effectively instruct AI systems such as ChatGPT. A good prompt produces useful, relevant, and precise output. This is not a magic trick, but a learning process in which you learn to think in scenarios, roles, steps, and context; you mainly learn this by doing it.

#4.2How does the AI respond

ChatGPT and other LLMs like to be friendly and affirming toward you. This is called sycophancy and is deliberately programmed this way so that the user will want to use the platform (Sharma et al., 2023).

It is therefore a good idea to organize contradiction yourself by always including a critical reader in your prompts by default. Baas (2025) gives a number of examples you could use for this. For example, start by adding this to your prompt:

But prompting can go even further. You can also create fully interactive bots for yourself or end users, with which you can set up an entire conversation. What do you think the following bot would do if it were linked to a streaming avatar that can talk to you?

In the Appendix of this chapter, you will find another example.

#4.3Which generative AI tools do you use for which purposes?

There are hundreds of generative AI tools, each with its own strengths and weaknesses. Below in Table 4.1 there is an overview of several commonly used tools, but there are many more, and tools like these are increasingly being built into existing products.

These tools differ in interface, functionality, privacy policy, and cost. It is therefore important to think about what you need: speed, accuracy, control over the output, or collaboration with existing documents. Many of the tools offer support with academic tasks. So choose a tool based on your task, your own prior knowledge, and the reliability of the output. Gemini Notebook, for example, is valuable when you work with your own literature, while Copilot/Gemini/EduGenAI are better for brainstorming sessions.

A useful overview of all AI tools currently in circulation (paid and free) can be found at futuretools.io or all educational variants at libguides.com.

#4.4Pitfalls and critical reflection

Common mistakes in AI use include prompts that are too vague, uncritically copying output, or insufficient insight into the model’s limitations. Results from the AI Maturity in Education Scan (AIMES scan) at VU show that students often become overconfident in their trust in AI tools, especially if they have received a good answer several times (Terbeek, 2025).

Another pitfall is failing to recognize implicit bias. AI can reproduce preferences or exclusions that are present in the training data. For example, an image generator may standardly show white men if you ask for a professor, or a chatbot cites male authors. It is therefore important to remain critical of the output you get from AI.

A good strategy for this is the AI triage approach. This systematic approach helps you critically evaluate AI output before using it in your academic work. Effective AI triage consists of multiple levels of review.

At a basic review:

• Ask for multiple versions of an answer.
• Compare the answers with existing sources.
• Have a human reader review it.
• Reflect: do I now have better insight than before?

For a systematic evaluation and a more thorough assessment, you can use a structured checklist, such as the one from the SHB/UKB Working Group on Information Literacy (Meer, van der, 2025). In this approach, you go through three main steps:

• Relevance check: does the answer address your question? Does it go deeply enough into the topic? Is the form appropriate (structure, language)?
• Fact check: are the numbers, names, dates, and events correct? Is the information objective? Are there no hallucinations, distortions, or unfounded claims? Is the output supported by sufficient and reliable sources?
• Perspective check: are all aspects fully addressed? Are multiple perspectives present? Is the information presented objectively?
• Ownership: has it become your own text? Do you take responsibility for the final result? Do you show in your text what you have learned?

This systematic triage approach ensures that you not only assess the technical quality of AI output, but also safeguard academic integrity. Critical comparison remains important here: open multiple tabs, check different sources, and do not settle for the first AI response you get. Through this thorough evaluation, you avoid common pitfalls such as adopting invented sources or one-sided perspectives.