Adaptive Prompt Engineering for LLM-Driven Feedback in Reflective Writing
Bachelor Thesis · BSc Wirtschaftsinformatik · Bern University of Applied Sciences (BFH Business School) Author: Cyril Schlup · Supervisors: Prof. Dr. Thiemo Wambsganss & Léane Wettstein · 2026
Context
Reflective writing is a core practice in higher education: it helps students turn their experiences into deeper learning. But the personalised feedback that makes reflection work does not scale, in large classes, individual guidance arrives late, vague, or not at all. Large Language Models promise a way out, since they can produce instant, individualised feedback. The catch lies in how they are trained: optimised to be maximally "helpful," they tend to rewrite passages, reinterpret the student's experience, and hand out generic praise. When the AI effectively co-writes the text, students stop feeling that it is theirs. This sense of psychological ownership is precisely what sustains motivation and engagement, so the very helpfulness of an LLM can quietly undermine the learning it is meant to support.
Goal & Tasks
The thesis investigates whether prompt engineering alone, working only on the input, without fine-tuning the model, can produce feedback that is both pedagogically effective and ownership-preserving. The concrete goal was to design prompt templates that achieve two things at once:
- reliably produce three structurally distinct feedback strategies, GAS (Giving-Answers, direct hints), PAS (Prompting-Answers, questions only), and EAS (Elaborating-Answers, interpret then question); and
- preserve the student's psychological ownership across all three.
Three research questions guided the work: how the strategies can be constructed through prompting (RQ1), how they influence psychological ownership (RQ2), and which transferable design principles can be derived (RQ3).
Methods
The project followed Design Science Research (Hevner's three-cycle model).
- Relevance cycle — grounding the work in practice: 92 expert feedback entries were classified into the three strategies, 10 student user stories were mapped onto three routes to ownership, and the existing MindBuddy prompt was analysed as a baseline.
- Design cycle — building the artifact: a modular seven-component prompt architecture was developed, with 12 templates (four per strategy). A dedicated ownership layer translated the three ownership routes into five concrete prompt constraints — subjunctive phrasing, text-specific anchoring, a replacement-text prohibition, interpretation respect, and a specific opening. 140 outputs were evaluated across two refinement rounds against four pass/fail criteria (strategy fidelity, ownership compliance, length compliance, content anchoring).
- Evaluation (FEDS framework) — combining artificial feasibility tests with a naturalistic user study: 12 participants used a browser-based React prototype running on Claude Sonnet, then rated an ownership questionnaire from. A post-hoc run on GPT-4 checked cross-model consistency.
Results
- All-pass rate improved from 21% to 54%; the five best templates reached 86% on held-out texts never seen during development.
- Ownership compliance rose from 34% to 85%, a 51-percentage-point gain.
- In the user study, all three strategies preserved ownership above the scale midpoint, with mean scores of 5.81–6.06 on a 7-point scale (PAS highest).
- Cross-model pass rates stayed within 5 percentage points between GPT-4 and Claude Sonnet.
- Central finding: without explicit constraints, two-thirds of outputs violated at least one ownership principle; with the designed templates, this dropped to 15%.
Beyond the artifact (12 templates on a reusable modular architecture), the thesis contributes three design principles, Strategy Differentiation, Ownership Preservation, and Length-Bounded Expression. The take-away: LLMs do not preserve students' ownership on their own, but carefully engineered prompts can make them do so without sacrificing feedback quality.