Key points are not available for this paper at this time.
arXiv priprint 2026-05-10 This paper analyzes an actual work process in which a user generated a manuscript and a natural-language IDE design-document package in a single long-running LLM session through natural-language instructions, criterion corrections, file inputs, error identification, and artifact regeneration requests. A key feature of this paper is that the main structure and body of the manuscript were generated through a sequence of natural-language instructions, criterion corrections, file inputs, error identification, and rewriting requests. The development environment in this case was a web-based ChatGPT Plus session. The user did not use a traditional programming language, manuscript-writing automation scripts, a dedicated natural-language IDE, API integration, or local development tools. Instead, using only an LLM as a natural-language runtime and an interactive web session, the user sequentially generated the title, abstract, table of contents, Sections 1 through 11, and Appendices A through F of the paper—not through a single complete prompt, but through a sequence of section-by-section output requests, criterion revisions, length adjustments, file inputs, error identification, and rewriting requests. For example, the body sections were generated through natural-language commands such as “1. Introduction,” “2. Related Concepts and Theoretical Background,” and “7. Operating Mechanism. Explain this section sufficiently.” The appendices were also generated step by step, from “Appendix A. Sample User Command Log” to “Appendix F. Example of a Natural-Language Source State.” In this process, the LLM generated each section and appendix based on the work narrative, criterion structure, constraints, artifact list, and current state cumulatively formed by the user. In this case, brainstorming, preliminary material generation, manuscript development, organization of supplementary materials, and preparation for public release were carried out within a continuous workflow on the scale of several hours. However, this study does not present this time information as a generalizable productivity metric. It treats it only as a supplementary observation showing how a manual natural-language development environment operated. The case shows that the accumulation of natural-language instructions and criterion corrections can lead to the formation of a manuscript, related appendices, and the structure of a design-document package. This paper interprets the process not as simple prompt engineering, but as an operational case of natural-language programming performed in a manual natural-language development environment. The user acted as a natural-language programmer by managing the work goal, document structure, criterion structure, direction of error correction, and artifact sequence. The LLM acted as a probabilistic natural-language runtime that generated artifacts under the given criterion structure. Accordingly, this paper has the character of a self-documenting case study: it explains the possibility of natural-language programming while also demonstrating that possibility through the actual process of generating the paper itself. In addition, to show the possibility of natural-language programming and to preserve the record of the work process, this paper generally preserves the structure of the generated work logs and artifacts without excessive refinement. At the same time, it allows edits that help readers understand the paper, including improvements in academic readability, sentence clarity, consistency of section structure, reference formatting, and reduction of redundant expressions. In other words, the paper preserves the original form generated through natural-language programming while also making the minimum readability improvements needed for a manuscript suitable for submission. This paper makes three contributions. First, it defines and analyzes user utterances that change the work state and criterion structure in an LLM session as natural-language programming operation logs. Second, by presenting a case in which an actual manuscript and design-document package were generated through section-by-section and appendix-by-appendix natural-language commands, it shows that natural-language programming can already be performed manually in current LLM environments. Third, from the limitations of this manual mode of operation, it derives the need for a natural-language IDE that supports work-narrative storage, criterion-structure management, current-state tracking, artifact-list management, session handoff, permission separation, and CSO/CVO-based verification.
Building similarity graph...
Analyzing shared references across papers
Loading...
Lee Hochul
Building similarity graph...
Analyzing shared references across papers
Loading...
Lee Hochul (Sun,) studied this question.
www.synapsesocial.com/papers/6a13e81d0e02ee3982d32c0d — DOI: https://doi.org/10.5281/zenodo.20360241
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: