Using AI to Guide AI Implementation in Science: An Exploratory Case Study in Human Genome Editing

Artificial intelligence is being adopted rapidly across scientific disciplines and is transforming research practice. Despite this, no general methodology exists for determining how AI should specifically be implemented in a given scientific context. Implementation decisions therefore rest with individual researchers, regardless of their expertise in AI or the scrutiny they apply to its risks. The harms of this approach are already measurable. Research has shown that current implementation practice has narrowed the scope of scientific topics and shifted research toward the automation of existing methods rather than the generation of new ideas. An improved methodology is urgently needed. This paper presents the first exploration of whether AI, already established as a powerful tool throughout scientific practice, can also be applied to the problem of its own implementation. The field of human embryonic genome editing was used as a case study to document a frontier model's capabilities, effective prompting strategies, and recurrent weaknesses on this task. An 18-step, three-phase prompting pipeline was developed iteratively with Claude Opus 4.6, and outputs were assessed qualitatively against published literature with manual verification of cited sources. Final outputs were verifiably novel and grounded in current literature, and verification steps within the pipeline identified and corrected errors introduced at earlier steps. Three recurrent weaknesses persisted despite prompt-level mitigation: instruction-following degradation on long prompts, overstatement of recommendation quality, and insufficient domain knowledge. These observations provide preliminary evidence that AI can serve as an effective tool for improving its own implementation in science, warranting sustained investigation across models and domains.