Year: 2026
Type: dissertation
Abstract: Over eighty percent of today's developers are using or planning to use AI coding assistant tools. However, the code produced by AI coding assistants often does not meet production-quality standards. Most commonly, these issues come from lack of layered architectures, hardcoded configuration values, insufficient input validation and discrepancies in how back end and front end types were implemented. These issues arise due to the fact that AI models do not retain specific project context over time. In order to address these issues, a full stack application template was created that embeds machine readable conventions into it to direct the behaviour of an AI. The template includes a root level file containing instructions for how an AI should behave, multiple domain specific rule files for directing backend and frontend behaviours, a spec driven process for generating code based upon a formally defined specification prior to running the code through the AI tool chain, and an automated pipeline that will ensure that changes made to the front-end client code are consistent with any changes made to the backend service definition. The effectiveness of this template was assessed by providing the same set of requirements to the same AI model in both a template-less environment and the template. In each case the AI produced fully operational applications. However, when scored against a checklist of 33 items covering aspects such as architecture, API contracts, validation, configuration and type safety, the template assisted output achieved 100%, while the template-less output achieved only 50%. Focus groups consisting of 7 software development professionals also noted that there were meaningful differences in terms of production readiness between the two sets of generated applications. They further rated both the instruction file concept and the spec driven concept as being valid (4.3/5). Of those surveyed, four indicated a willingness to adopt the template. Two others stated that they did not believe it would reduce their overall workload complexity sufficiently (3.6/5). Overall, the study demonstrates that creating project conventions in machine readable files can improve non-functional code quality. The most significant improvements occur in areas related to architecture, configuration management and API consistency. Future research directions will include reducing the complexity associated with implementing this approach.
Source: Theseus (Ammattikorkeakoulujen)
Author: Nguyen, Bao
Institution: JAMK University of Applied Sciences
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162175330
Cites:
Cited by:
Related to: