From Prompt Engineering to Epistemic Prompting: Prompt Trajectories as AI-Mediated Problem Framing in Science Education

Technical Summary: From Prompt Engineering to Epistemic Prompting

Problem Statement
Current approaches to Large Language Model (LLM) interaction in STEM education predominantly treat "prompt engineering" as a technical competence aimed at optimizing output quality, accuracy, or format. While this downstream evaluation logic is necessary for tool utility, it is insufficient for learning. The paper identifies a critical gap: existing literature fails to theorize how sequential learner turns function as disciplinary problem-framing moves within an accumulating conversational context. Furthermore, there is a risk that fluent model outputs may conceal inappropriate models, unstated assumptions, or mismatches between representations, leading to reduced cognitive effort and the acceptance of unsupported claims. The central problem is the lack of a framework that interprets prompt trajectories not merely as iterative refinements for better answers, but as epistemic practices where learners articulate, regulate, and evaluate how a knowledge task is constituted.

Methodology
This article is a conceptual model paper (Jaakkola, 2020) that synthesizes existing literature from four primary research strands: prompt engineering/AI literacy, conversational tutoring, metacognition/self-regulation, and epistemological framing/multiple representations. Rather than presenting new empirical data, the author constructs a theoretical framework by mapping the "unresolved issues" in these adjacent fields. The methodology involves:

1. Synthesizing Definitions: Integrating "epistemological framing" (the situated interpretation of activity, knowledge, and action) with the mechanics of human-LLM dialogue.
2. Operationalizing Concepts: Developing a functional vocabulary (Table 2) to code learner turns based on their impact on the "epistemic organization" of a problem.
3. Modeling Interaction: Proposing a recursive "Framing-Prompting Loop" that visualizes the relationship between prompt function, accumulated context, disciplinary standards, and learner uptake.

Key Contributions
The paper introduces the concept of Epistemic Prompting, defined as the deliberate construction and revision of prompts to articulate, regulate, and evaluate how a knowledge task is framed. Its primary criterion is "epistemic adequacy"—whether the interaction makes the objects, relations, assumptions, and distribution of labor available for scrutiny—rather than mere output quality.

The core contributions include:

• The Prompt Trajectory: Shifting the unit of analysis from the isolated prompt or final answer to the entire sequence of prompts, model responses, learner uptake, and disciplinary checks.
• The Framing-Prompting Loop: A multi-turn model where an initial prompt establishes a provisional macro-frame (selecting problems, representations, assumptions, and work distribution), and subsequent turns maintain, specify, challenge, repair, transform, or reframe that organization.
• Functional Taxonomy: A set of operational indicators for coding learner turns (Table 2), distinguishing between:
  • Establishing (defining the problem and role).
  • Maintaining (preserving the current model).
  • Specifying (making implicit elements explicit).
  • Challenging (questioning evidence or validity).
  • Repairing (diagnosing inconsistencies while preserving warranted elements).
  • Transforming Representation (shifting between modes, e.g., graph to equation, for disciplinary purpose).
  • Reframing (introducing new models or criteria).
  • Closing/Synthesizing (requesting validation and justified closure).
• Concept-to-Decision Warranting: A specific criterion requiring that prompts connect concepts to concrete decisions (e.g., how a concept constrains a calculation or validation criterion), moving beyond simple definition or answer-seeking.

Results and Claims
As a conceptual paper, the "results" are the proposed framework and its theoretical implications for STEM instruction and assessment:

• Distinction from Prompt Engineering: Unlike conventional prompt engineering which seeks to optimize output, epistemic prompting treats the prompt as a public trace of the learner's epistemic agency. A prompt is educationally productive not because it yields a perfect answer, but because it exposes a misconception, a representational conflict, or a missing validation criterion.
• Regulated Delegation: The framework advocates for "regulated delegation," where the model contributes representations, calculations, or critiques, but the learner retains normative responsibility for framing, validation, and determining when a conclusion is warranted.
• Instructional Implications: Instruction should move beyond "templates" for good outputs. Instead, tasks should require students to formulate prompts that explicitly state the disciplinary context, role distribution, methodological path, and validation criteria. Students should be trained to revise prompts to challenge assumptions or shift representations, rather than simply rephrasing for clarity.
• Assessment Implications: Assessment should evaluate the prompt trajectory (process evidence) rather than just the final prompt or answer. Rubrics should examine the adequacy of initial framing, the quality of uptake, the function of follow-ups, and the justification of closure, looking for patterns of unproductive answer-seeking versus productive epistemic regulation.

Significance and Limitations
The paper claims that epistemic prompting extends AI literacy by connecting operational competence with responsibility for inquiry. It argues that in an era of accessible generative AI, the decisive competence is maintaining epistemic control over the inquiry process.

The author explicitly notes the paper's modesty regarding its claims:

• No Empirical Evidence: The article proposes a framework but does not provide empirical evidence that epistemic prompting instruction improves understanding, transfer, or self-regulation.
• Equity Concerns: The framework acknowledges that elaborate natural-language prompting may privilege students with strong academic language or prior disciplinary knowledge. It warns against conflating linguistic fluency with epistemic competence and suggests that instruction must provide multiple entry points (e.g., sketches, code, spoken dialogue).
• Future Directions: The paper calls for future research to test the framework across disciplines, examine the relationship between trajectory types (e.g., challenge/repair vs. answer-seeking) and independent problem-solving, and validate the coding categories against independent measures of conceptual understanding.