Background. As LLMs become a routine “interaction layer” inside software engineering toolchains, teams increasingly rely on prompts, retrieval scope, and lightweight checks to steer behavior without access to model internals, while still needing accountability, reproducibility, and cost awareness across the Software Development Life Cycle (SDLC). Motivating scenario. Consider an organization where multiple LLM-backed tools are invoked throughout development (requirements critique, design reviews, code review, testing, maintenance) across heterogeneous repositories, evolving policies, and sometimes different jurisdictions. In such a setting, the practical governance problem is not a single task prompt, but the need to make interaction-time controls reviewable and replayable: what policy and context were bound at invocation, what sources were retrievable, what checks ran, what trade-offs were accepted, and what evidence supports the output. Aim and scope. This paper proposes an evidence-informed roadmap to 2030 for Ethical Prompt Engineering (EPE), scoped strictly to interaction-time. It frames prompts, policy/context bindings, retrieval scope, and checks as versioned, auditable artifacts integrated into existing SE workflows, while making explicit what EPE cannot guarantee and when escalation beyond the prompt layer is required. Method. Adopting a design-science stance, the roadmap is built via a focused rapid review that consolidates convergent evidence from prompting surveys, LLM-for-SE syntheses, and governance-oriented sources, extending the authors’ earlier SE2030 workshop version with broader evidence and a capability-oriented synthesis plus reporting guidance. Results and contributions. The roadmap delivers: (i) a challenge map centered on policy/context binding, provenance for review, robustness and transfer across models and repositories, sustainability of safeguards, and human oversight; (ii) harmonized evaluation constructs and a reporting kit that co-reports utility with ethical compliance, robustness to prompt perturbations, traceability from constraints to output spans/tool calls, and sustainability overhead with full configuration disclosure; and (iii) a reference capability model that groups policy/context handling, prompt planning/patterning, provenance/audit, and evaluation/reporting, mapped across SDLC activities and distilled into actionable agenda cards to 2030. Boundaries and implications. Within its remit, EPE can constrain outputs, scope retrieval to vetted artifacts, require citations, and capture provenance; it cannot remove entrenched representational harms or unintended memorization, so the roadmap treats escalation to model/data or organizational governance as a first-class requirement, alongside risk-based and cost/energy-transparent adoption.
Migliarini et al. (Thu,) studied this question.