Large language models are now used throughout literature review workflows, including summarization, extraction support, comparison across studies, and manuscript preparation. Existing disclosure norms have responded mainly by asking authors to identify whether generative AI was used, which tool was used, and for what general purpose. These requirements are necessary, but they are insufficient when prompts and iterative model interactions contribute to the interpretive work of literature synthesis. This article develops a threshold-based framework for prompt disclosure in qualitative and mixed-methods research. Drawing on construct validity, qualitative audit-trail traditions, qualitative evidence synthesis, review-reporting standards, and recent AI disclosure guidance, it argues that prompts should be treated as methodological traces when they shape cross-paper relationships, explanatory claims, causal hierarchies, typologies, metaphors, or adopted interpretive frames. In such cases, disclosure should extend beyond the prompt text itself to include the interaction path, supplied context, model and settings where available, decision points, human evaluation criteria, and the rationale by which model-generated interpretations were accepted, rejected, or revised. Lighter disclosure remains appropriate for grammar editing, formatting, translation used only for comprehension, unused brainstorming, and descriptive summary that does not enter synthesis. The proposed checklist links disclosure burden to interpretive consequence rather than to AI use as such. It treats prompt disclosure as an auditability practice that strengthens construct-validity warrants in AI-mediated knowledge production, without promising exact reproducibility.