Generative AI (GenAI) models, especially large language models (LLMs), are increasingly deployed in education for tasks such as item generation, automated scoring, and feedback delivery. The quality of these outputs depends not only on the model but also on how users craft prompts to steer model behavior. Existing prompt engineering relies on ad‑hoc trial‑and‑error, which introduces validity threats like construct misrepresentation, construct‑irrelevant variance, and limited reproducibility.\
Evidence‑Centered Design (ECD) is a measurement framework that structures valid inferences from observed performance. We map ECD’s three core models—the student model, the evidence model, and the task model—onto prompt design components: the construct to be measured, the observable evidence, and the task description. This yields an ECD‑based prompt engineering workflow that (1) articulates the target construct in the prompt, (2) defines observable evidence and designs prompts to elicit it, and (3) builds a task model ensuring the prompt aligns with assessment goals.\
From this mapping we derive five design principles: construct alignment, evidence observability, task context consistency, response interpretability, and fairness control. We then conduct an empirical study on automated essay scoring, contrasting ECD‑crafted prompts with state‑of‑the‑art baselines. Results show that ECD prompts improve scoring accuracy, reliability, and fairness, confirming that treating prompting as assessment design rather than informal instruction enhances AI‑generated educational outputs.\
In summary, Evidence‑Centered Design offers a principled, verifiable framework for prompt engineering, strengthening the validity, reliability, and equity of AI‑driven educational tools. Review