I let Gemma pick my top 3 tasks, and the rest of the list can wait

My to-do list is not a plan. It is a museum of things I felt guilty about once.

So for Hacktoberfest Week 3 ("Slow and Steady"), I built the smallest thing that fixes
that: sloth-triage, a ~70-line Bun CLI that hands your task list to an open-weight
model and asks the only question that matters — what are three things actually worth
doing today? Everything else is explicitly allowed to wait.

$ bun run index.js tasks.txt

🦥 sloth-triage (gemma4:cloud) read 9 tasks and chose 3

1. Finish the quarterly report — due tomorrow at 9am.
2. Renew the domain — it expires Friday.
3. Reply to the landlord — a leaking tap gets worse.

Skip the conference recording and the photo library. The downloads folder can wait forever.

That is the whole product. No accounts, no dashboard, no streaks, no "productivity
system". It reads a text file, makes one request to a model on my machine, and prints
a short plan. Then it stops.

The build

There is nothing to install:

{
  "name": "sloth-triage",
  "type": "module",
  "scripts": { "start": "bun run index.js tasks.txt" }
}

The core is one prompt and one fetch to an OpenAI-compatible endpoint:

const prompt = `You are a very lazy but sensible assistant. Below is the user's task list.

Pick the ${top} tasks that genuinely deserve attention today. Prefer deadlines and things
that get worse if ignored. Then write a short "lazy plan": order the tasks, one line each,
and say what can be deliberately skipped or deferred. Be calm and brief.

Tasks:
${tasks.map((task) => `- ${task}`).join("\n")}`;

const res = await fetch(`${baseUrl}/chat/completions`, {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    model,
    messages: [{ role: "user", content: prompt }],
    temperature: 0.3,
  }),
});

SLOTH_MODEL and SLOTH_BASE_URL are the only configuration. The default is Ollama on
127.0.0.1:11434, running Gemma. Lines can be plain text or - [ ] checkboxes; --top=N
changes how many survivors you get.

Why open matters here

A task list is one of the most personal files on a laptop. It contains health
appointments, money problems, names of people I owe replies to. Sending that to a closed
API so it can rank nine bullets is a bad trade.

With an open-weight model served by Ollama, the list stays on the machine:

• No account and no API key. The tool works the same on a plane.
• No meter running. A 1B model answering one short prompt costs effectively nothing.
• The model is a setting, not a vendor. Swap Gemma for Llama or Qwen with one env var, and the prompt is right there in the file if you want to change how it thinks.
• The whole thing is auditable. Two files and no dependencies. You can read every line before it reads your diary.

That last point is the real one. "Open source AI" is not just a license on weights. It is
being able to see the prompt, see the request, and know there is no hidden second call.
A closed product cannot offer that, no matter how good its summary is.

What I would fix next

• Small models sometimes confuse "urgent" with "loud". The prompt asks for reasons so the bad calls are at least visible.
• It does not know my calendar. A real version would pass in the next two days of events.
• I tested with Gemma through Ollama; the code expects any local OpenAI-compatible server.

The project is deliberately boring: one file, one prompt, one output. That is the point.
A tool that helps you do less should not itself become another thing to maintain.

Repo: https://github.com/ItzmeJan/sloth-triage

Built for the Hacktoberfest Open-Source AI Challenge: Week 3.