Persona-first prompts: how FitGuide used lightweight personas to reduce coach hallucinations

AI · 5 min read

FitGuide, an AI-powered fitness coaching startup, had been relying on a single generic system prompt for its conversational coach. Users reported that advice could be inconsistent, and unattended hallucinations—like suggesting unsafe exercises—undermined credibility. The product team decided to embed lightweight persona scaffolds and content constraints into the prompt layer rather than retraining models.

Designers created a persona-selector UI where users choose from a small set of evidence-based coaching styles: Motivational, Evidence-first, and Clinical. Each persona sets boundaries and style guidelines for the model; for example, Clinical limits exercises for common injuries and requires citations when making medical claims. The UI also exposes a “safety footer” explaining when professional medical advice is recommended.

A/B tests showed persona-selected sessions had 45% fewer hallucinations (as flagged by a safety classifier) and a 30% rise in user-reported trust. Users appreciated clarity: seeing the coaching style ahead of time reduced surprise and increased perceived reliability. The product also introduced a lightweight feedback loop where users could flag questionable advice, which trained a moderation layer that improved over time.

For AI product design, the FitGuide case shows how lightweight UX controls—persona selection, explicit constraints, and transparent safety signals—can make generative systems safer and more useful without expensive model retraining. Startups can often get big behavioral improvements by designing better prompt UX and feedback loops.