This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
TOUCH-WOOD: An AI That Wants You to
Stop Using It 🌿
An idea for the Hacktoberfest Open-Source AI Challenge — Week 1: Touch Grass.
«What if we built an AI whose success was measured by how quickly people could put their phones away?»
A Note About This Submission
As a college student, balancing my academic schedule and examinations made it difficult for me to complete the implementation during this challenge week. I couldn't build the working prototype within the available time, so this submission focuses on the problem, product concept, proposed technical architecture, and implementation plan.
I want to be transparent about the current status rather than present an idea as a completed product.
What I Want to Build
We have applications designed to keep us scrolling, algorithms competing for our attention, and AI assistants that can answer almost any question.
But here's something I've been thinking about:
What if AI could help us spend less time using technology?
We often open our phones for a few seconds and end up spending an hour scrolling. Students spend long days in front of screens, remote workers can go hours without leaving their desks, and even our breaks have become another opportunity to consume digital content.
I want to explore a different approach.
TOUCH-WOOD is a proposed AI-powered application that turns a person's available free time into a small, personalized, real-world activity.
Instead of recommending another video, generating an endless conversation, or asking users to spend more time in an app, TOUCH-WOOD would give them one practical mission to complete in the physical world.
It could encourage someone to take a short walk, notice details in a familiar place, explore their surroundings, move their body, or take a mindful break.
The idea is simple:
Let AI create the experience. Let people live it. Then let the screen disappear.
How Would TOUCH-WOOD Work?
Imagine a student who has spent several hours studying for an examination. They feel mentally exhausted and automatically reach for their phone.
Instead of opening another social media application, they open TOUCH-WOOD.
The application asks:
"What do you need right now?"
They can choose from six intentions:
- Clear My Head: Take a short break and reset.
- Move My Body: Complete a simple walking or movement activity.
- Explore: Discover something unfamiliar nearby.
- Notice: Pay attention to things usually ignored.
- Connect: Find a small opportunity for safe, real-world social interaction.
- Surprise Me: Let AI choose a suitable activity.
Next, they select how much time they have: 5, 10, 20, or 30 minutes.
Based on their preferences, available time, accessibility needs, and chosen environment, the AI generates one practical mission.
Example Mission: Rediscover the Ordinary
Duration: 15 minutes
Objective: Experience a familiar place differently.
- Choose a safe walking route that you already know.
- Put your phone away and avoid unnecessary notifications.
- Notice three things you have never paid attention to before.
- At the end, recall one detail that surprised you.
Mission complete: You experienced your surroundings more consciously. No photograph, public post, or digital proof is required.
Someone else might receive a completely different mission, such as finding a quiet outdoor spot, observing the shapes of clouds, or taking a short movement break indoors when going outside is not practical.
The application should adapt to the person rather than assume everyone has access to a park, a walkable neighbourhood, or the same physical abilities.
The Most Important Feature: Phone Down Mode
This is the heart of TOUCH-WOOD.
Most applications are designed to keep users engaged. TOUCH-WOOD would deliberately reduce interaction once the user has received their mission.
After starting an activity, the screen becomes minimal:
YOUR MISSION HAS STARTED.
You don't need me right now.
Go outside, explore, move, or simply pay attention to the world around you.
When the activity is finished, the user can return if they want to record a reflection.
No endless feed. No pressure to maintain a streak. No mandatory photographs. No leaderboard encouraging people to compete over their free time.
The application should not turn spending less time on a screen into another digital addiction.
The goal is not to replace one habit with another. It is to make technology useful and then unnecessary.
What Happens When You Return?
After completing a mission, TOUCH-WOOD could offer an optional reflection.
For example:
"Welcome back. How was your experience?"
The user can select a response, write a short note, or simply finish without recording anything.
An AI model could optionally turn a reflection into a short, encouraging summary.
For example, a user might write:
"I took a walk near my home and noticed a small garden I had never paid attention to."
The application might respond:
"You found something new in a familiar place. Sometimes a change in attention is enough to make an ordinary day feel different."
With the user's permission, the application could save completed missions and reflections to provide more varied suggestions in the future.
The emphasis would remain on the real-world experience, not on collecting personal data.
Who Is TOUCH-WOOD For?
I would initially design the application for:
- Students who spend long hours studying or working on screens.
- People who want to reduce unnecessary scrolling.
- Remote workers who rarely leave their workspaces.
- Beginners who want simple, approachable ways to explore their surroundings.
- Anyone who wants a short break without planning an entire outdoor activity.
The experience should be accessible and flexible. A user should not be required to walk a particular distance, visit a specific location, interact with strangers, or share their location to benefit from the application.
If going outdoors is unsafe or impractical, the application should offer an appropriate indoor alternative.
How I Plan to Use Open-Source AI
I don't want to add AI merely because this is an AI challenge. I want the model to perform a meaningful task that would otherwise require manually writing and maintaining a large collection of activities.
My proposed approach is to use an open-weight model from the Gemma family for personalized mission generation.
The model could receive structured information such as:
- Available time.
- The user's selected intention.
- Preferred activity intensity.
- Accessibility requirements.
- A general environment category, such as urban, park, or indoor.
- Previously completed missions, if the user chooses to save them.
The model would then generate a single activity using a defined output format.
Example of a Proposed AI Response
{
"title": "Notice the Ordinary",
"duration_minutes": 10,
"intention": "mindful_exploration",
"steps": [
"Take a short, safe walk.",
"Notice three details around you.",
"Recall one detail before returning."
],
"phone_down": true
}
This is an illustrative example of the proposed response format, not output from a working integration.
The backend would validate the generated response before showing it to the user. It would also handle malformed responses, generation failures, and safe fallback missions.
The AI should not invent live weather conditions, claim that an unverified location is safe, or recommend activities that conflict with a user's accessibility needs.
The initial version would not require continuous GPS tracking. Users could select their environment manually, keeping the system simpler and more privacy-conscious.
Proposed Technical Architecture
The following is my proposed architecture for a future implementation. These technologies have not yet been integrated into a working prototype.
- Frontend — Next.js, TypeScript and Tailwind CSS
The frontend would provide a clean, mobile-first interface with five main screens:
- Home and mission preferences.
- AI-generated mission.
- Phone Down Mode.
- Optional post-mission reflection.
- Personal activity history.
A progressive web app could make the application convenient to access on mobile devices.
- Open-Weight AI — Gemma
Gemma would handle mission generation and potentially short, optional reflections.
I would start with a small, well-defined generation task rather than building a complicated autonomous agent.
The final model and inference provider would be selected after evaluating response quality, latency, cost, and the challenge's requirements.
- Backend — API and Validation Layer
A lightweight backend would:
- Receive the user's selected preferences.
- Validate the input.
- Construct a structured model request.
- Call the chosen inference service.
- Validate the returned mission.
- Return the mission to the frontend.
The backend would also handle rate limiting, API errors, and fallback activities.
- Database — MongoDB Atlas
If persistent storage proves useful, MongoDB Atlas could store optional user preferences, completed mission metadata, and reflections.
The core mission experience should not depend on collecting sensitive personal information. Location history would not be necessary for the initial version.
- Deployment and Monitoring — Render and Sentry
Render is a proposed platform for deploying the application or backend, depending on the final architecture.
Sentry could help diagnose application errors and identify failures during testing.
These integrations would be evaluated and implemented only if they meaningfully support the project and satisfy the relevant challenge rules.
Why Does Open Innovation Matter?
Open innovation matters because it gives developers more flexibility in how AI systems are built, evaluated, and adapted.
An open-weight model can provide opportunities to experiment with different inference environments, customize model instructions, and evaluate alternatives without making every aspect of the application dependent on one proprietary model API.
For TOUCH-WOOD, this flexibility could help answer practical questions:
- Can a smaller model generate useful missions reliably?
- How much context is actually necessary?
- Can structured outputs reduce unpredictable responses?
- Can the model be replaced without redesigning the entire application?
- Can inference costs and latency be reduced without compromising usefulness?
I would also distinguish between using an open-weight model through a hosted inference service and running that model locally. These are different deployment choices, and the final implementation would depend on the available resources and challenge requirements.
The larger idea is that open innovation allows developers to experiment with the role AI plays in people's lives, not simply add more AI-generated content to existing digital habits.
What Makes TOUCH-WOOD Different?
TOUCH-WOOD is not intended to be another hiking planner, nature-identification application, or chatbot that recommends outdoor activities.
Its central principle is that reduced screen time can be a successful outcome.
That principle influences the proposed design:
- Missions are short and actionable.
- Users do not need to continue chatting with the model.
- Completion does not require uploading evidence.
- Reflection is optional.
- Personalization should improve usefulness rather than increase screen time.
- Safety and accessibility should take priority over completing a mission.
- The application should remain useful even if the user does not save any personal data.
The most important success metric would not be the number of messages exchanged with the AI.
It would be whether users find the missions practical and whether the experience helps them engage in activities beyond their screens.
Proposed Development Roadmap
If I continue developing TOUCH-WOOD, I would build it in the following phases.
Phase 1 — Minimum Viable Product
Create the responsive interface, preference selection, mission screen, and Phone Down Mode.
Phase 2 — AI Integration
Integrate the selected Gemma model, implement structured generation, validate responses, and create fallback missions.
Phase 3 — Reflections and History
Add optional reflections, simple activity history, and privacy controls.
Phase 4 — Testing and Deployment
Test mobile usability, mission quality, accessibility, error handling, and the complete user journey. Deploy the application and document the actual results.
Phase 5 — Real-World Evaluation
Conduct a small, voluntary usability test to understand whether the missions are practical, accessible, and genuinely useful.
These are planned development phases, not completed milestones.
Current Status and What I Learned
TOUCH-WOOD is currently a concept proposal. I have not completed the implementation, deployed a working application, or tested the proposed AI integration.
My college schedule and examinations prevented me from building the prototype during this challenge week.
Although I would have preferred to submit a working application, developing this concept made me think more carefully about what makes technology useful.
A technically impressive application is not necessarily a good product if it encourages the same behaviour it claims to solve.
For TOUCH-WOOD, the challenge is not only generating a good mission. It is designing an experience that helps people act in the real world without creating another reason to stay on their phones.
That is the problem I would like to explore through this project.
Final Thoughts
We often measure AI by how much content it generates, how long it can hold a conversation, or how many tasks it can automate.
I want to explore a different measure.
Can AI help someone take a break, notice their surroundings, move a little, or experience an ordinary day differently—and then let them get on with their life?
TOUCH-WOOD is my proposed answer to that question.
The best interaction might be the one that ends with the user putting their phone away.
Submitted as a concept proposal for the Hacktoberfest Open-Source AI Challenge — Week 1: Touch Grass. Implementation remains pending.
<\THANK YOU FOR YOUR TIME/>
Top comments (0)