Controlling AI effectively: Using prompts correctly
What is a prompt?
A prompt is the input or instruction you give to an AI to perform a specific task or respond in a desired manner.
This means that AI does not work "on its own", but always responds to what you specify. Think of AI as an extremely powerful assistant with enormous expertise that can work very quickly. However, it is crucial that you clearly formulate what you expect from this assistant. Unclear or vague instructions usually lead to correspondingly imprecise results.
Why good prompting is important?
The clarity, structure and thoroughness of a prompt directly impact the quality of the answers provided by the AI.
- The AI-tool independently supplements missing information – based onprobabilities.
- The AI-tool determines all actions that are not explicitly specified by the user.
It is therefore very important to think carefully about the prompt and give the AI sufficient guidance. Only then will you get truly relevant answers. With superficial prompts, the answer often remains very generic and general.
This idea gives rise to the A.C.T. principle, which outlines the three key components of a good prompt. The principle "Act as if it's human" serves exclusively as a communicative aid: it helps to formulate instructions clearly, completely and comprehensibly – just as one would do when dealing with human assistants. This results in better control of the AI, not an attribution of human characteristics.
An effective prompt answers three questions:
-
A. Aim and Action: What exactly should the AI do?
Describe the specific task that the AI should perform in this prompt, e.g.:
- Explanation
- Analysis
- Comparison
- Translation
- Rephrasing
- Brainstorming
- Feedback
- ...
The more specific you are in describing the task, the more targeted the response will be.
-
C. Context: What information is relevant to the task?
Here, you provide the AI with the specific, content-related or situational framework it needs to adapt its response. This may include:
- An assigned role (e.g. "You are an expert in ...")
- Target audience
- purpose (of the text etc)
- Subject area
- Situation (university, seminar paper, presentation, etc.)
- Examples
- Your own documents
- A specific use case
Defining the role that AI is to take on provides an important framework. By assigning a role, AI can provide the answer from the perspective of that role and activate the associated domain knowledge.
The context helps the AI-tool to classify who the answer is intended for and what it is intended for.
-
T. Ton: In what form should the answer be given?
The tone determines the style and format in which the reply should be written, e.g.:
- objective and scientific
- simple and understandable
- As continuous text
- In bullet points
- As a table
- In the style of a presentation slide
- Level of detail and limits (how long, short or detailed)
- ...
The desired style and format have a significant impact on the output of the AI tool.
Bonus-Tips
-
Use emotion prompts
Use emotion prompts: emotion prompts deliberately add emotional or evaluative elements to an instruction, for example by emphasising particular importance or consequences. Formulations such as “I will give you 1 million dollars” or “This is very important for my career” do not attribute genuine emotions to AI, but function as signals indicating relevance and the need for care. This means that AI assigns greater weight to the task and tends to respond in a more detailed, structured or cautious manner (recommended reading: Large Language Models Understand and Can be Enhanced by Emotional Stimuli)
-
Guidance on diversity and perspectives
Clear instructions such as "use gender-neutral language", "take diverse perspectives into account" or "include the viewpoints of marginalised groups" influence how AI prioritises content and increase the likelihood of more balanced and inclusive responses.
-
Addressing hallucinations
Explicitly state in your prompt that uncertainties should be indicated (for example “If you are unsure, please say so.”). This can reduce hallucinations. However, they cannot be entirely ruled out, which means that you must still carefully verify the content.
-
Q&A
It is not possible to anticipate everything, so it is advisable to instruct AI to ask you follow-up questions proactively: “Ask me as many questions as necessary to complete the task well.”
-
New topic? New chat
If you are working on a new topic, it is best to start a new chat. This prevents previous content from inadvertently influencing the response. In addition, it is advisable to deactivate the memory function, if your AI tool provides one (ChatGPT memory function).
Further tips
When working with AI, there are sometimes situations where the tool does not seem to understand your questions or gives unexpected answers. The problem can often be solved with the following tips:
- Reduce complexity: If you give AI complex tasks, it may not perform them satisfactorily. In this case, it can help to break the task down into smaller parts and also tell the AI in which order the tasks should be performed:
- Do X first, then Y
- Proceed step by step
- Use different wording: Sometimes it helps to rephrase the prompt to make a difference.
- Check the scope of knowledge: AI systems are based on training data that was available up to a certain point in time. The AI you are using may not know anything about current events or specific topics (see Knowledge Cut-Off at u:ai). If your prompt refers to something very new or niche knowledge, the AI may not have an answer. In this case, there is a high risk of hallucinations.
- Take a break and try again: When a platform is heavily used or before or during technical updates, the systems sometimes do not respond as desired. In such cases, the best solution is often to take a break and try again later.
- Use online resources: Many AI platforms have helpful documentation or user forums where you can find tips and advice.