Hey there,

Usually I write about AI for work — building agents, shipping product, the stuff non-technical builders are figuring out.

Today, I want to share a personal AI use case.

I'm in Japan right now, and I've been running a small experiment that's made me reflect a bit on the implications of AI on our daily lives.

Travel content about Japan is everywhere right now. Walk around a touristy area of Tokyo and you'll inevitably see a long line of people waiting for a "viral" restaurant. But that's not what I'm here for. What interests me most about Japan is shokunin-run places — the small coffee shops, listening bars, and restaurants where a single owner has mastered their craft and runs the place solo.

You don't usually see these places on the "must-visit" lists that ChatGPT or Claude pull from.

Taste is being averaged across millions of people, and the average doesn't match anyone in particular.

Instead of asking AI for recommendations the way you'd ask a search engine, I spent the upfront time telling it about me. Specific neighborhoods I've been drawn to in other cities. Cities I've loved and cities I haven't. The vibe of a place that makes me want to stay versus leave. I packaged all of it into Claude Skills — reusable context files the model loads every time I ask it something.

Then I asked the usual travel questions. Where to stay. Where to eat. What to do on a Saturday afternoon. I actually had it plan entire train journeys, then almost fully book them using Claude in my browser.

It didn't send me to the viral places. It sent me to neighborhoods that matched the ones I'd loved in other cities. It surfaced restaurants whose descriptions read like places I'd actually pick. The hit rate wasn't perfect — it's still pulling from blogs and indexed content, so the floor is whatever the internet has already written down — but it was meaningfully closer to my taste than any list I could have found by searching.

It's also still imperfect in the obvious ways. It can over-index to something you said, it can make mistakes. It's not a replacement for walking around and finding your best spots by accident.

If you've heard the term "context engineering," this is basically it — making your preferences, constraints, and taste legible to the model.

Once you see this concept, it shows up everywhere. Anywhere AI is currently giving you the same answer it gives everyone else, the fix is the same: you don't only need to ask better questions, you should provide better context.

People have genuinely mixed opinions on AI right now. Some see efficiency, some see risk, some worry it'll make us think less or push everyone toward the same average answer. Most of us hold a few of those at once. I think about it a lot too. But this trip was a small reminder that the story isn't fixed yet. Used well, AI doesn't have to flatten taste — it might be one of the few tools that actually pushes back against the algorithms already doing that.

I'm putting the skills I built on GitHub — genericized so you can fork them, swap in your own taste, and use them for your next trip. Link here.

Curious what you find when the algorithm stops choosing for you.

Enjoy!

Shaalin