This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
What I Built
Soft Start turns an unfiltered pile of tasks into one small first step, a short optional plan, and a list of things that can wait. Pick a pace that matches your energy, paste the list you have been avoiding, and let a local model make it easier to begin.
It is for anyone who spends more energy deciding where to start than doing the first small thing. The plan is deliberately modest: stopping after the first step still counts.
Demo
There is no public deployment yet. To try it locally, install Ollama, download a model such as llama3.2:3b, then serve the project folder with python3 -m http.server 4173 and open http://localhost:4173. If the browser blocks the local model connection, follow the OLLAMA_ORIGINS note in the README.
Code
The project is a static HTML, CSS, and JavaScript app with no build step. Its source is in index.html, styles.css, and app.js; setup instructions are in README.md. A public repository link is not available yet.
How I Built It
Soft Start calls Ollama's local /api/tags endpoint to list installed models and /api/chat to generate a structured plan. The user chooses the model, and the prompt is visible and editable in app.js. Different pace settings change the guidance sent to the model, so the user can decide how ambitious the plan should be.
The app itself has no server, account, API key, analytics, external font requests, or cloud AI dependency. It sends the task list only to the Ollama service running on the same device. Once Ollama and a model are installed, inference can run offline without a per-request API charge.
The implementation was syntax-checked, but I could not exercise a live generation in this environment because Ollama and a downloaded model were not available.
Why Does Open Innovation Matter?
Task lists can contain personal details. Local inference keeps them off a hosted AI service, and offline use means the tool can still work without an internet connection after setup. Open-weight models also make the choice reversible: install a different model or adjust the prompt to change the assistant's behavior, without changing providers or rewriting the app around a closed API.
That flexibility matters here more than chasing a single benchmark. A small model that runs on your own machine can turn a stuck moment into a first step, with no subscription or usage meter attached.
Top comments (0)