ReConGen: Requirements-Constrained Code Generation

Name of participant: Jonathan Ullrich

Project’s name: ReConGen: Automated integration of requirements and architectural decisions into code generation in complex software projects

Project description:

Ensuring architectural consistency remains a major challenge in AI-assisted software development. While large language models (LLMs) already support developers in generating code, they frequently ignore or inconsistently apply project-specific architectural decisions and constraints. As a result, generated code often violates established design principles, requiring significant manual correction and limiting practical efficiency gains.

Existing approaches attempt to address this issue by providing architectural guidelines as additional context, for example in the form of structured documentation, such as AGENTS.md files or similar artifacts. However, these approaches do not reliably enforce compliance and lack the ability to encompass a broader scope of architectural constraints. Due to limitations in context handling (“lost in the middle” problem) and the tendency of coding agents to hallucinate or overlook constraints, important architectural decisions are often not consistently reflected in the generated code.

The ReConGen project addresses this gap by focusing on the systematic enforcement of architecture conformity in AI-based code generation. The goal is to transform architectural decisions and requirements into representations that can be treated as binding constraints during the generation process, rather than optional contextual information.

The project investigates how relevant project knowledge (i.e., architectural decisions and requirements) can be retrieved from existing documentation and translated into enforceable conditions for code generation. This includes identifying which constraints are critical for maintaining architectural integrity, as well as developing methods to represent them in a way that can be consistently applied by LLM-based systems.

ReConGen contributes to a better understanding of how architectural compliance can be assured in AI-assisted development. From an industry perspective, it enables more reliable integration of code generation tools into existing development processes by ensuring compliance with existing architectural decisions. In the long term, this can lead to more maintainable systems, reduced development effort, and increased trust in AI-supported software engineering.

Software Campus Partner: Fraunhofer IESE and DATEV

Implementation period: 01.03.2026 – 28.02.2028