Itinera — AI travel itineraries you can actually edit

Why I built it

I travel a lot, and writing day-by-day itineraries by hand is tedious. So I built one for myself: punch in from, to, dates, budget, and how I'm getting there — Gemini drafts the plan, and I save it as a trip I can edit later.

The eventual scope is bigger than the current build — nearby stay suggestions, restaurants, and commute availability per day — but the goal of this first cut was to get the core planning flow right: search → AI plan → editable trip.

How it works

1. Auth. Better Auth handles email/password + Google OAuth. The username plugin gives each user a clean profile slug.
2. Search. The /api/trips/search route shapes the form input, calls Gemini with a few-shot prompt, and validates the response shape before persisting.
3. Persist. Each search is stored as a SavedSearch. Promote one to a Trip and you get day-by-day editable rows (ItineraryDay), a trip-status state machine (PLANNED → UPCOMING → IN_PROGRESS → COMPLETED), and post-trip review + rating.

Choices worth calling out

Better Auth. I picked it because it was new and I wanted to try it — a straightforward learning-driven choice rather than a deep tool comparison. The Prisma adapter and the username plugin saved me a chunk of model code I'd otherwise have written by hand.

Few-shot prompting for Gemini. Instead of trusting a "please return JSON" instruction alone, I include 1–2 example outputs in the prompt. The model picks up the shape from the examples and returns far more consistent day-by-day structure than instruction-only prompts do. The response is then schema-validated before it ever touches the database.

What broke

The hardest part wasn't the AI — it was the Prisma ↔ Better Auth integration. Better Auth ships a Prisma adapter, but the docs largely assume you're starting from a clean slate. Retrofitting onto an evolving schema meant working through field-name expectations, migration ordering, and session-handling quirks more carefully than I'd planned for. None of it was one-line.

Status

A proof of concept, built over 2–3 weeks of after-work evenings. The core flow works end-to-end on the live demo, but it's not production-ready yet. Next up: the nearby-stay / restaurant / commute layer the original scope called for, and tightening the Gemini prompt so the occasional surreal day suggestion stops slipping through.

Stack

Next.js 15 · React 19 · TypeScript · Prisma 6 + Postgres · Better Auth · Google Gemini · shadcn/ui · Tailwind v4 · Vercel Blob

Live demo · Source