Aril

Why is finding a gift so draining?

A birthday, an anniversary, a milestone — gift time arrives and the first stop is usually a search engine or a listicle. The result is always the same: suggestions broad enough to apply to anyone, personal enough to fit no one. Even people who know the recipient well tend to freeze on 'what do I get them?' The problem isn't personalisation — it's personalisation at the moment you need it, without the legwork.

Design questions, not search

My first instinct was a search interface: user types keywords, model matches. I tried that direction and dropped it quickly — this isn't a catalogue problem, it's an interpretation problem. The right move was to let the model infer what the user couldn't articulate themselves.

Instead I built a small sequential question flow: relationship to the recipient, rough budget, a few interest signals. The prompt treats those answers as context and produces both concrete and reachable gift ideas. Refreshing the list is possible with a single tap, producing a new set.

On the Supabase side I kept the schema minimal: user sessions and saved lists. I didn't go after a caching layer for suggestions — each request is short and latency is acceptable as-is.

Why Claude, not GPT?

GPT-4o would have been technically capable here too. The deciding difference for Aril was predictability and tone. Claude's responses read flatter, carry less 'AI voice' — for gift recommendations that matters, because users want a suggestion from someone with taste, not from a system.

Expo was the obvious mobile layer: single codebase for iOS and Android, and the managed workflow meant store submission friction stayed low.

What shipped

A working iOS and Android app that produces a personalised gift list in under a minute from a few short answers. Users can refresh the list or save it. The interface is deliberately sparse — the experience should feel effortless, not cognitively loaded.

What I'd see differently now

This project showed most clearly how directly prompt engineering shapes output quality. The first version was functional but generic; small changes to tone and question ordering produced noticeably better results. In a next version I'd structure suggestion categories more deliberately from the start.