SciDraft: A Custom GPT for Paragraph-Level, Human-in-the-Loop Scientific Manuscript Drafting

Overview SciDraft is a Custom GPT specialized for paragraph-level, human-in-the-loop scientific manuscript drafting. This record contains a Technical Note (preprint) describing the design principles and implementation of SciDraft, as well as characteristics observed during its use. It does not include quantitative performance evaluation or comparative validation through user studies. Background and Approach Scientific manuscript writing is a time-consuming process that requires maintaining logical coherence, appropriate formatting, and careful structuring of information. While large language models (LLMs) have recently been explored as tools for assisting academic writing, existing tools often generate entire manuscripts or sections in a single output, making it difficult for researchers to preserve fine-grained control over content and reasoning. To address this issue, we developed SciDraft, a custom GPT designed to support scientific manuscript drafting through an interactive and paragraph-level process. SciDraft guides users through a stepwise procedure that begins with defining key elements such as the research question and structure, and iteratively generates and refines each paragraph with continuous researcher confirmation and revision. This workflow is designed to align with established principles for effective human–AI interaction, ensuring that researchers remain responsible for the final content while benefiting from AI-assisted drafting. Although this Technical Note does not include quantitative evaluation, SciDraft provides a reproducible design framework for integrating generative AI into scientific writing in a controlled and transparent manner. Key points - Paragraph-level drafting workflow with explicit human checkpoints (approval/revision)- Starts from existing research materials (e.g., grant proposals, progress slides) to preserve the researcher's conceptual framework- Designed as a collaborative agent supporting researchers rather than an autonomous writing system What this record contains - SciDraft_manuscript_Zenodo_20260116.pdf (manuscript) Project page / access Japanese version: https://note.com/daichi_konno/n/nc941e5eacfb5English version: https://note.com/daichi_konno/n/n376fa4bd4946?app_launch=false