AI when it earns the click

A lot of AI-in-productivity products feel like they want to run when you are not looking. Background chats, automatic re-rankings, quiet rewrites of things you typed. The result is a product that is hard to predict and harder to trust.

We picked a different shape. AI in Mini Me OS only runs when you press a button. There are three buttons. That is the whole surface.

The three buttons

Smart add — on Today and Capture. You type a sentence. You press the button. The AI proposes a structured list of tasks, goals, projects, routines or notes. A preview opens. You confirm or cancel. The preview can be edited line by line.

Break into steps — inside the task editor. You press the button. The AI proposes 3 to 6 sub-tasks for the task you are looking at. A preview opens with checkboxes. You tick the ones you want; everything else is discarded.

Explain today's plan — on the Assistant page. You press the preset. The AI reads a compact snapshot of your Today screen (counts, top task, attention items, workload status) and writes a 60-to-120-word paragraph in plain English. No save, no side effect — just reading material.

Three buttons. Previews everywhere. Nothing else in the app calls AI. The regular Add and Save buttons never do.

What gets sent, exactly

We are precise about this on the AI disclosure page (/legal/ai). The headline is: only the specific text you submit to the feature you triggered.

• Smart add sends the sentence you typed.
• Break into steps sends the task title, the description if you wrote one, and the name of the project if the task belongs to one.
• Explain today's plan sends counts (how many tasks, how many overdue), the title of the top task, a short list of attention items, and the workload status.

That is it. The AI never sees your full task list, your email, your goals as a list, or your calendar contents.

The fallback

When the operator has not configured a Gemini API key, all three features fall back to a built-in rule-based version that runs locally on our server. Smart add still splits sentences (using regex and heuristics). Break into steps still proposes 3 generic steps. Explain today's plan still composes a paragraph from templates. The buttons never break.

The UI tells the user honestly which one ran — AI suggestion or Rule-based — with a small badge in every preview.

Why this shape

Two reasons.

One — predictability. A product where AI runs only on a button click is a product where the user can build an accurate model of when it is doing what. That is the whole point of a calm operating system.

Two — honesty. If we ran a background re-ranker every time you typed, the engine's reasons would shift under you and we would have to explain why. Keeping AI out of the silent paths means the engine's reasons are stable, knowable, and the same on every device.

Three buttons. Previews everywhere. We will keep it boring on purpose.