From Sketches to Specs: AI-Assisted Lightweight Metamodeling for Spec-Driven Development
Abstract
Sketching is a natural medium for early-phase software activities. However, preserving the resulting information beyond just taking photos is difficult and laborious. A decade ago, the FlexiSketch approach addressed this problem by combining free-form sketching with lightweight metamodeling capabilities, thus enabling a smooth evolution from sketches to semi-formal artifacts. The advent of large language models (LLMs) and specification-driven development (SDD) make it worthwhile to revisit the FlexiSketch approach. LLMs can support tasks such as shape recognition or metamodel inference that were previously out of reach. SDD uses a GenAI pipeline that generates code from structured specifications. The structured models produced by an AIenhanced FlexiSketch approach could be a natural fit for such specifications. We envisage an AI-assisted sketch-based workflow which lets stakeholders engage in early-phase modeling with creative freedom, while producing structured artifacts that are suitable for feeding an SDD pipeline. In this position paper, we explore our vision: we identify points in the sketch-to-code pipeline where AI assistance for a FlexiSketch style approach is beneficial, propose design principles for that assistance, and outline a research agenda.