Letting AI Describe Diagrams as JSON Instead of Drawing Them Yields Better Results

A developer tested the same 44-node architecture diagram twice using an AI model — once requesting raw SVG output with hand-placed coordinates, and once asking the model to output structured JSON for a dedicated layout engine to render. While modern AI models handle labeled boxes adequately, the SVG approach consistently produced tangled edges and unstable layouts because the model predicts tokens sequentially rather than solving global spatial constraints. The developer argues this is a structural limitation, not a capability gap, since routing connectors across complex graphs is precisely the problem tools like the ELK layout engine are built to solve. Their alternative pipeline, called Glyphic, has the model emit plain JSON describing nodes and edges, which is then validated against a strict schema before ELK computes positions and Rust-based resvg rasterizes the final image — with no headless browser required. The key advantage is that JSON output is diffable, editable, and consistently reproducible, whereas regenerating SVG from scratch on every change produces a different result each time.