Jonathan Gordon defines a complete design-code roundtrip as editing in both directions while preserving fidelity and the origin of each element. Design and engineering have long worked toward different constraints, and a model that can interpret design intent as well as generate code appears to offer a way to connect them. His experience testing these workflows instead exposed gaps between convincing visual output and dependable implementation.
Jonathan Gordon describes stopping an AI coding session after noticing an innerHTML operation he considered risky. Looking more closely at generated code led him to investigate claims that code-to-design and design-to-code tools had completed the roundtrip. He reports trying five tool configurations and encountering losses such as broken bindings or changes that survived on only one side, rather than presenting those tests as an exhaustive comparison of every available tool.
In a live demonstration, Jonathan Gordon places generated code beside a design representation, makes a form change and runs ReWeaver AI to identify discrepancies. The findings span design tokens, code quality, performance and accessibility. One example flags a missing ARIA live region, which he explains would prevent a screen-reader announcement, and shows a proposed fix removing that finding.
Jonathan Gordon argues for deterministic checks around probabilistic generation, with clear provenance linking design elements to their source code. A proposed correction should either satisfy a known rule or be refused and left for human judgment. He also describes a small repeated-generation experiment, but the talk does not provide enough methodological detail to treat its quality percentages as a general benchmark.
Jonathan Gordon distinguishes human control over changes from simply placing a reviewer inside an agent-led process. In the workflow he presents, a person can apply, undo or ignore a finding, while design consistency and accessibility remain explicit review dimensions. The broader lesson is to inspect what changes across successive generations and catch divergence before it reaches a merge.
Watch on YouTube
