Applying SHAPR in AI-Assisted Research Software Development: Lessons Learnt from Building a Share Trading System
For researchers using generative AI in software development, this paper offers practical guidance and good practices to mitigate issues of continuity and traceability, though the findings are based on a single case study and are incremental.
This paper presents a case study of applying the SHAPR framework to develop a modular share trading system, demonstrating that continuous documentation updates and structured practices (e.g., contracts, source-of-truth documents) help maintain project coherence and knowledge continuity in AI-assisted research software development. Five recurring lessons were identified, including that contracts stabilize AI-assisted coding and that environment setup contributes to knowledge generation.
Generative AI is changing how research software is developed, but rapid AI-assisted development can weaken continuity, traceability, and methodological clarity. SHAPR (Solo, Human-centred, AI-assisted PRactice) was proposed as a framework for structuring AI-assisted research software development. This paper presents a documented case of applying SHAPR to the development of a modular share trading system. From the outset, the project adopted a SHAPR-informed working configuration that shaped how interaction, implementation, and documentation were organised. Across iterative development cycles, the project generated a structured evidence base including reflection notes, development cycle review notes, source-of-truth documents, contracts, quick captures, workflow notes, and evolving code artefacts. The case showed that continuous documentation updates, supported by quick capture and AI-assisted refinement, helped maintain organised and usable project knowledge throughout development. Five recurring lessons were identified: contracts stabilised AI-assisted coding, a maintained source-of-truth layer improved coherence, cycle-boundary snapshots strengthened continuity, code and documentation co-evolved through quick capture and iterative refinement, and environment setup itself contributed to knowledge generation. The case also illustrates a practical SHAPR operating configuration in which a ChatGPT Project and cycle-specific chats supported interaction, reasoning, summarisation, and coding collaboration, PyCharm supported artefact implementation, and Obsidian supported external working memory, structured documentation, reflection, continuity, and repository-oriented note organisation, while remaining consistent with SHAPR's tool-agnostic principle. The paper contributes practical guidance and good practices for researchers conducting AI-assisted research software development.