I built a local AI that tells me to go outside (Gemma + Ollama, works offline)

TanviMalewar — Sun, 11 Oct 2026 19:15:26 +0000

What I built

Touch Grass Prompter is a tiny web app that runs entirely on my laptop. You pick the weather, how many minutes you have, and your mood. A local Gemma model gives you one short outdoor mini-mission. Then you close the laptop and go do it.

The screen is meant to be the shortest part of the experience: about 4 seconds to get a mission, then 15 minutes outside.

Why open-weight AI mattered here

I chose an open-weight model on purpose, and it changed what the app could be:

Works with no internet. After the one-time model download, I turned off Wi-Fi and it kept working. That matters for something meant to be used outdoors.
Private by default. My mood, weather, and time never leave my laptop. There is no server and no API key.
Free to run. No per-request cost, so I can click the button as many times as I like.
Swappable. The model is one constant in server.js. Changing gemma3:1b to another model is a one-line edit.

How it works

A plain HTML page collects weather, minutes, and mood.
A small Node/Express server builds a prompt and sends it to Ollama's local API.
Gemma returns one short, concrete mission.
The page shows the mission and how long it took to generate locally.

Stack: Gemma (via Ollama), Node.js, Express, vanilla HTML/CSS/JS.

The prompt asks for one mission, three short sentences, and nothing that needs a phone or screen.

Taking it outside

I picked "light rain", "15 minutes", and "tired", and the app gave me my mission in 3.7 seconds, fully local.

[WRITE 2 TO 3 SENTENCES HERE: where you went, what you did, how it felt.]

What worked and what didn't

Latency was about 3.7 seconds per mission with the 1B model, which is fast enough that I never felt like I was waiting on a screen.
The small model sometimes invents things it can't know. In my test it said "I've spotted a patch of wildflowers", but it can't see where I am. The missions work as inspiration, not as facts. A bigger Gemma would likely do better, and swapping it in takes one line.

What I'd improve next

Add the user's real local weather as input instead of a dropdown.
Try a larger Gemma and compare quality against speed.
Make it a phone-friendly offline app.

Code

GitHub: https://github.com/TanviMalewar/touch-grass-prompter

Entering for Best Use of Gemma.This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

Prize Categories

Best Use of Gemma: the whole app runs on Gemma, served locally through Ollama. It generates every mission offline on my laptop, and I can swap the model by changing one line.

DEV Community: TanviMalewar