@inproceedings{holzman-etal-2026-clinical,
    title = "Clinical Prompt Engineering: Encoding Clinical Knowledge into {AI} Training Simulations - A Crisis Deployment Case Study",
    author = "Holzman, Yuval  and
      Rafaeli, Eshkol  and
      Elyoseph, Zohar  and
      Haber, Yuval  and
      Yirmiya, Karen  and
      Linkovski, Omer  and
      Elyoseph, Tal  and
      Refoua, Elad",
    editor = "Zirikly, Aya  and
      Bar, Kfir  and
      MacAvaney, Sean  and
      Ireland, Molly  and
      Ophir, Yaakov  and
      Atzil-Slonim, Dana  and
      Varadarajan, Vasudha  and
      Bedrick, Steven  and
      Desmet, Bart",
    booktitle = "Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology ({CLP}sych 2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.clpsych-1.3/",
    pages = "32--42",
    ISBN = "979-8-89176-421-7",
    abstract = "When large language models simulate patients or clients, they tend to produce cooperative dialogue, premature emotional insight, and rapid resolution. These defaults undermine clinical training, where the pedagogical value lies in sustained difficulty. We describe Clinical Prompt Engineering (CPE), a methodology developed by a multidisciplinary team of clinician-researchers and prompt engineering experts within the [ProjectName] project. CPE encodes clinical knowledge directly into prompt design: each simulated character is constructed through layered psychological profiles, explicit contingency rules linking interactional events to internal states, and enforced non-linear emotional trajectories that resist the model{'}s pull toward resolution. The methodology has been applied across several clinical training simulations involving over 300 participants in formal studies and iterative pilot phases. Each simulated character is embedded within a multi-agent training environment that provides real-time reflective guidance during the interaction and structured, clinically informed feedback afterward. We illustrate the approach through Talking with Lia, a Hebrew-language simulation in which parents practice responding to a seven-year-old child during repeated missile alerts and forced sheltering. The simulation was deployed within the first week of an acute security crisis in Israel in Winter 2026. Of 132 sessions initiated organically through professional networks, 42 were completed; qualitative feedback emphasized the simulation{'}s difficulty as pedagogically meaningful. Because CPE operates at the level of prompt design, it can be developed by clinician-researcher teams and adapted to new populations, developmental stages, and crisis contexts, potentially extending access to expert-informed training beyond the settings where such expertise is typically available. Where much computational work in clinical psychology has focused on classifying mental health states from text, CPE addresses a complementary task: whether clinicians can respond effectively to those states as they shift in real time. The next step is testing whether the skills practiced in simulation transfer to real interactions."
}