How an LLM-Powered Design Assistant Reduced Iteration Time at Retail UX Startup ShelfSense

AI · 5 min read

ShelfSense, a retail analytics startup, struggled with slow handoffs between designers and copywriters: every component needed tuned microcopy, CTA variants, and localized strings for three markets. The team introduced an LLM-powered design assistant embedded in Figma that generated copy variants, suggested label hierarchy, and flagged contrast issues.

To avoid common LLM pitfalls they fine-tuned the assistant on their brand voice and a corpus of approved microcopy, and built strict guardrails — the assistant returns only three variants per prompt, includes a confidence score, and tags suggestions that require legal review (pricing, claims). Generated content flows into a lightweight review queue where a designer approves or edits before export to the dev repo.

Within two months iteration time on new components dropped by 47%, with designers reporting faster ideation and fewer stale copy rounds. Metrics showed the review acceptance rate for assistant suggestions stabilized at 68%, with the remainder edited for tone or legal compliance. Accessibility flagging caught 24 contrast issues before they hit QA.

ShelfSense's approach demonstrates that LLMs can be productive in design pipelines when constrained by domain-specific data and human-in-the-loop checks. The team emphasized that the assistant augmented creative work rather than replaced it: designers retained final authority, using the tool to accelerate mundane but essential tasks.