tatami

Designing the AI coaching feature really highlighted a big decision in how I built the prompt architecture. If you use a "stateless" weekly summary, the output eventually gets generic, and users stop trusting the AI after a week or two. To fix that, I added a persistent memory layer in Supabase. By injecting the last three weeks of recommendations into each API call as structured context, the AI stopped just pattern-matching raw data and actually started providing mentorship that felt consistent over time.

The other big hurdle was getting the tone right. At first, telling the AI to be "direct and honest" made the feedback sound almost punitive, which is a disaster for a hobbyist audience. I had to add explicit negative constraints - like "never guilt-trip" and "never lecture" - while reframing the AI’s persona from an authoritative coach to more of an experienced training partner.