This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: Interview mit KI Forscherin Prof. Kim über Prompt Engineering.

Professor Inhwa Kim gave her students at Oregon State University a fixed amount of time and one task: produce the best retail layout for a clothing company's store, using generative AI. No restriction on which tool, and no template for the prompt. A goal and a deadline, and nothing else.

Afterwards she went through the prompts systematically, separating the ones that had produced the strongest results from the ones that had produced the weakest. The pattern was not accidental: the good results shared four properties, and the weak ones lacked all four.

Choosing the tool is already a decision

The best results came from students who did not open whichever system was nearest. They weighed up which provider suited this particular task before starting, because no model is equally good at everything. Understanding that means making a decision before the first prompt that shapes the entire output.

Context down to the detail

Vague inputs produced generic results: layouts that would have fitted any store anywhere. The strongest work came from students who specified the target customer, the brand style, the size of the space, the logic of the product range, the atmosphere they wanted, even the paths customers would walk.

The more context the prompt carried, the less the model had to guess, and the less often it guessed wrong.

Treating the output as a draft

Students who accepted the first suggestion they were given did worse across the board. The strongest results came from those who treated what came back as a draft, a starting point for revision and not a finished product.

That matches how research on prompt literacy defines the skill, with iterative refinement built into it. Quality comes from iteration, and the first output rarely delivers it.

Structure inside the prompt itself

The order in which the information arrived made a visible difference. Putting context before task before format specification produced better results than packing everything into one unstructured paragraph. The model responds to structure, which is why a structured question tends to return a structured answer.

The part Kim had not expected

What she emphasized most in conversation was this: the students who did best were not the most technically adept. They were the ones who knew most clearly what they wanted before typing the first prompt.

It fits what her earlier research on communication with retail robots already suggested. In human-AI interaction the decisive factor is the empathetic quality of the human communication, which has to adapt itself to the limits of a machine, and not the capability of the system. Ask vaguely and you get vague answers, because the system can only work with what it is given.

Why this applies well outside a classroom

Kim's experiment ran in a university setting, with students generating a retail layout, and the principles that come out of it hold for any professional use.

A consultant prompting for a market strategy faces the same questions as Kim's students. How much context do I supply? Which model do I choose? Do I take the first output or iterate? And above all: do I know what I want before I ask?

Research on prompt literacy, the term for the ability to use AI systems precisely, reflectively and critically, supplies the framework. Kim's experiment makes the framework tangible. The finding goes past the fact that better prompts produce better results. It names which specific properties make the difference, and locates the decisive variable in the deliberate, responsible approach of the person, not in the tool.

On 9 April 2026 we ran our first prompt engineering workshop at WU Vienna, hosted by Ladies that Lead, and built Kim's findings into the program: how to construct a prompt step by step, and how to check reliably at the end whether the output holds. The motivation and curiosity in the room were the most encouraging part of it, and there is a great deal of room for AI teaching at universities.

Sources

Grouped by the section they support.

Opening

  • No external source. The classroom experiment and its evaluation come from the author's interview with Inhwa Kim.

Choosing the tool is already a decision

  • No external source. The observation is Kim's, from the interview.

Context down to the detail

  • No external source. The observation is Kim's, from the interview.

Treating the output as a draft

  • Hwang, Lee and Shin (2023), What is prompt literacy? An exploratory study of language learners' development of new literacy skill using generative AI, arXiv:2311.05373. Supports iterative refinement as part of prompt literacy.

Structure inside the prompt itself

  • No external source. The observation is Kim's, from the interview.

The part Kim had not expected

  • Kim (2025), From Adoption to Optimization of AI-Powered Retail Service Robots: Consumer Switching and Communication Effectiveness, doctoral dissertation, University of Tennessee, Knoxville, trace.tennessee.edu. Supports her earlier research on how retail robots communicate. The interpretation is Kim's and the author's.

Why this applies well outside a classroom

  • Hwang, Lee and Shin (2023). Supports the concept of prompt literacy. The transfer to professional use is the author's own.
The link has been copied!