The Failure of the Design-Code Roundtrip
Despite the promise of AI-driven development, a true, lossless bidirectional roundtrip between design (Figma/Sketch) and code remains unsolved. Current workflows are inherently lossy, often resulting in dropped bindings or design changes that fail to propagate to the codebase. Jonathan Gordon defines a successful roundtrip as a full loop with persistent provenance, where every UI element can be traced back to the specific line of code that generated it. Without this, teams suffer from "drift," which Gordon identifies as the new form of technical debt that accumulates silently as AI agents generate code without oversight.
Deterministic Guardrails vs. Pure Model Output
Pure LLM output for UI generation is probabilistic, not deterministic. In experiments, pure model output typically starts at roughly 30% fidelity and degrades over time. To combat this, ReWeaver AI introduces a harness that scans both code and design canvases across nine dimensions, including accessibility, performance, and design token consistency.
Key insights from this approach include:
- Accessibility as a First-Class Citizen: Models often generate inaccessible code (e.g., missing ARIA live regions). Deterministic guardrails can detect these failures and suggest fixes, rather than relying on the model to "know" accessibility standards.
- Human-in-Control: Rather than letting agents operate autonomously, the human remains in control. The system provides a "show drift" feature that identifies mismatches, allowing the developer to accept or refuse specific fixes. This ensures that the final 10% of quality—which requires human judgment—is preserved.
- Governance: By treating AI-generated code as a managed asset, teams can govern the output before it is merged, preventing the accumulation of unverified code patterns like insecure
innerHTMLinjections.