This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.
I built Field Break, a deliberately small AI tool whose success condition is that you stop using it.
You tell it four things:
It returns one micro-adventure that can start almost immediately, plus a final instruction telling you when to lock the screen.
That last part is the product idea. Most AI apps optimize for more conversation. Field Break optimizes for the shortest useful interaction possible.
The 36-second demo is here:
https://raw.githubusercontent.com/ondmindmanagement-hub/field-break/main/demo/field-break-demo.mp4
A live Apertus 1.5 70B test was also captured during the challenge. Given 20 minutes, medium energy, neighbourhood streets and a small park, the model produced a “Neighbourhood Pocket Adventure” and ended with:
Turn off phone and let adventure unfold offline
The exact prompt and result are preserved in the repository so the model evidence is inspectable rather than described from memory.
Hacktoberfest Open-Source AI Challenge — Week 1: Touch Grass
Field Break turns a small amount of free time into one simple outdoor micro-adventure, then explicitly tells the user to put the screen away.
The app is designed for the open-weight Apertus 1.5 model through an OpenAI-compatible endpoint. The provider/model are environment-configurable, so the project can switch to another open-weight deployment without changing the UI.
Run with real model inference by setting:
OPEN_MODEL_API_KEY OPEN_MODEL_API_URL=https://api.publicai.co/v1/chat/completions OPEN_MODEL_NAME=swiss-ai/apertus-v1.5-8b
Without a key, the prototype runs a clearly labelled deterministic demo-policy; it never pretends that fallback output is model inference.
The planning layer is not locked to one proprietary model or API. An open-weight model can be self-hosted, swapped, audited, or moved closer to the user's data. For a tiny tool whose purpose is to get you away from the screen, that simplicity matters.
Field…
Repository: https://github.com/ondmindmanagement-hub/field-break
The project is intentionally dependency-light: a Python standard-library server, a small browser UI, three unit tests, and an OpenAI-compatible adapter for an open-weight model endpoint.
The open model at the center is Apertus 1.5, with the default configuration pointing to swiss-ai/apertus-v1.5-8b.
The model receives only the context the user provides. The system prompt explicitly tells it not to invent live weather, trail closures, local conditions, or medical claims. It must prefer something local, simple, low-cost and reversible.
The output is structured JSON:
screen_exit instruction.
That structure matters because I did not want another chat interface that can drift into an endless conversation. Field Break asks for one decision and then gets out of the way.
The provider and model are environment-configurable:
OPEN_MODEL_API_KEY=...
OPEN_MODEL_API_URL=https://api.publicai.co/v1/chat/completions
OPEN_MODEL_NAME=swiss-ai/apertus-v1.5-8b
If no model key is configured, the app switches to an explicitly labelled demo-policy. It never claims that deterministic fallback output came from the model.
I also added tests for JSON extraction, output normalization and the demo planner. All three are passing.
For this project, open innovation is not a decorative technology choice.
A closed planning API would make the smallest part of the product — deciding what to do outside — dependent on one vendor. With an open-weight model, the inference layer can be moved, swapped or self-hosted without changing the experience.
That becomes especially interesting for a tool whose philosophy is less cloud, less screen, less dependency.
The project can use a hosted Apertus endpoint today, but the interface is deliberately compatible with a future local or privately hosted deployment. The user interface does not need to know which provider is behind it.
Open weights also make the model choice inspectable. The project is not pretending that “AI” is a magic black box; the model family, prompt, evidence and fallback behavior are all visible in the repo.
The challenge theme pushed me to reverse the normal metric.
Instead of “How many messages can the user send?”, I asked:
How quickly can the software become unnecessary?
That changed the design:
The best output is not the most impressive paragraph. It is the one that makes someone say “okay” and leave the desk.
I am entering the overall Hacktoberfest Open-Source AI Challenge Week 1 category.
I did not add a partner category just to increase eligibility; the project only claims technologies it actually uses.
Field Break is small on purpose. The screen should be the shortest part of the experience.