Improving LLM-assisted code generation through the use of architectural documents and implementation plans
Proceedings of the 3rd International Workshop on Designing Software, pp 23-30
28 Jun 2026
In this paper, we investigate LLM-assisted code generation using architectural documentation as a first-class input. Building on prior work that guides LLMs through Attribute-Driven Design (ADD), we propose a workflow in which the LLM synthesizes an implementation plan from requirements and an architecture document, and then generates code with human oversight. Using a ReAct-style agent as a case study, we perform three experiments and evaluate architectural conformance and functional correctness of a primary use case. In addition we capture size and modularity metrics using the DV8 tool. Our results show that incorporating architectural documentation substantially improves conformance and that adding an explicit implementation plan further enhances functional correctness and modularity.
1
- Improving LLM-assisted code generation through the use of architectural documents and implementation plans
- Humberto Cervantes - Universidad Autónoma MetropolitanaRick Kazman - University of Hawaii SystemYuanfang Cai (Corresponding Author) - Drexel University
- Proceedings of the 3rd International Workshop on Designing Software, pp 23-30
- Designing '26: 3rd International Workshop on Designing Software (Rio de Janeiro, Brazil, 12 Apr 2026–18 Apr 2026)
- ACM Conferences
- Association for Computing Machinery (ACM)
- 8
- Conference proceeding
- English
- Computer Science
- 991022193998204721