Field Break: an open-weight AI that tells you to close the screen

Overview

Field Break is a deliberately small AI tool whose primary success condition is that you stop using it. The application takes four inputs from the user: how many minutes they have, their energy level, the kind of outdoor space nearby, and one constraint (e.g., "stay close to home" or "no car"). In return, it produces one micro-adventure that can begin almost immediately, along with a final instruction telling the user when to lock the screen. This last component is the core product idea-most AI applications optimize for extended conversation, whereas Field Break optimizes for the shortest useful interaction possible.

How It Works

Input Parameters

The model is prompted with four specific pieces of context:

• Minutes available - the time budget for the activity
• Energy level - a self-reported measure of physical/mental readiness
• Outdoor space type - describes the immediate environment (neighbourhood streets, parks, etc.)
• Constraint - a limiting factor such as "stay close to home" or "no car"

Output Format

The response is structured as JSON containing five fields:

• title - a brief description of the adventure
• number_of_minutes - the allocated duration
• short_plan - a concise actionable itinerary
• what_to_bring - items needed for the activity
• screen_exit_instruction - explicit guidance to disengage from the device

This strict structure ensures the tool stays focused on delivering value and exiting, rather than drifting into an endless conversation.

Technical Implementation

Model and Deployment

The core model is Apertus 1.5, running via an OpenAI-compatible endpoint. The default configuration points to swiss-ai/apertus-v1.5-8b. Because the provider and model are environment-configurable, the project can swap in another open-weight deployment without modifying the user interface.

For real model inference, the following environment variables control the setup:

OPEN_MODEL_API_KEY=
OPEN_MODEL_API_URL=https://api.publicai.co/v1/chat/completions
OPEN_MODEL_NAME=swiss-ai/apertus-v1.5-8b

When no API key is configured, the prototype falls back to a clearly labelled demo-policy. Importantly, the system never claims that deterministic fallback output originated from the model-it is explicitly labeled as a demo policy.

Dependency Lightness

The entire application is intentionally minimal:

• A Python standard-library server
• A small browser-based UI
• Three unit tests
• An OpenAI-compatible adapter for the open-weight model endpoint

This lightweight stack supports the philosophy that the smallest part of the product-the planning logic for what to do outside the app-should not depend on any single vendor.

Why Open Innovation Matters

Open innovation is central to Field Break's design. Rather than locking the planning layer to a proprietary API, the architecture allows the inference layer to be moved, swapped, or self-hosted without altering the user experience. This approach is particularly meaningful for a tool whose stated goals include being less cloud-dependent, less screen-reliant, and less tied to external providers.

Because the model family, prompt, evidence, and fallback behavior are all visible in the repository, the project avoids treating "AI" as a black box. The transparency extends to the model itself-users can inspect exactly what the system knows and how it decides to respond.

Build Process and Testing

The project was created during the Hacktoberfest Open-Source AI Challenge Week 1. Key build characteristics include:

• New project created within the challenge window
• Public GitHub repository: https://github.com/ondmindmanagement-hub/field-break
• Three unit tests covering JSON extraction, output normalization, and the demo planner-all passing
• Live Apertus 1.5 70B evidence captured from Public AI / CSCS infrastructure
• Short demo video available at https://raw.githubusercontent.com/ondmindmanagement-hub/field-break/main/demo/field-break-demo.mp4

Lessons Learned

The challenge theme forced a reversal of conventional metrics. Instead of measuring "how many messages can the user send," the focus shifted to "how quickly can the software become unnecessary?" This change reshaped the design around:

• No persistent feed
• No history requirements
• No engagement loops
• Exactly one plan
• Exactly one exit instruction

The optimal outcome is not the most impressive paragraph but the one that prompts a user to say "okay" and leave their desk-a tangible reduction in screen dependency.

---

Repository: https://github.com/ondmindmanagement-hub/field-break

Model: Apertus 1.5 (swiss-ai/apertus-v1.5-8b)