When the LLM took the wireframe: Inside Quill's generative UX experiment

AI · 6 min read

Quill's team built a plugin that translated a short product brief into a mid-fidelity wireframe and interaction spec using a fine-tuned LLM plus a rules layer to ensure accessibility constraints. The goal was not to replace designers but to accelerate the 0→1 exploratory phase. Designers were given the same brief and could either start from scratch (control) or use the generated wireframe as a starting point (experimental).

In sessions timed across ten designers and five briefs, the LLM-assisted group produced usable prototypes 48% faster on average. Designers reported that the generator helped avoid blank-page anxiety and surfaced edge-case flows they might have missed. However, quality checks showed subtle consistency issues: inconsistent spacing tokens, occasional unclear labeling, and one recommendation that violated platform guidelines. The rules layer caught many but not all problems.

Quill pivoted from a fully autonomous generator to a mixed-initiative pattern: the LLM proposes multiple variants, highlights accessibility assumptions, and attaches a “confidence” score tied to rule violations. The takeaway for product teams: generative models accelerate early exploration but require guardrails, explicit human checkpoints, and a focus on composability rather than end-product generation.