Skip to main content

Attention is the scarce thing: the me app

An agent that works all day floods you, so I built an app that gives me one decision at a time, ranked and paced through my day, and I steer it by saying "slower" or "break".

Attention is the scarce thing. If everything an agent makes reaches me, the agent has moved its work onto me. So the app's job isn't features. It's deciding what reaches me, when, and in what shape, and making every answer one tap.

  • Do the work first, then ask. A card is the end of the work, not the start. The agent prepares it until my part is a tap: a short summary, the recommended choice first, a deadline, and how much of my time it will take.
  • If I'm choosing something I'll look at, show it. When I'm picking a design, a page or a style, the card carries a preview for each option.
  • Everything goes in one queue, and the queue decides. The agent doesn't ping me. It queues. A small function ranks items by deadline, value and my effort, and the pacer sends one card at a time with breathing room in between, never early in the morning or late at night. Only a short written list of real emergencies skips the line.
  • I steer it like I'd talk to a person. I can say "break", "slower" or "more", and there is no settings screen.
  • One answer, run once. A tap in the app or in chat goes through the same function, and a repeat tap returns the original result.

Why rules became code: the pacing started as prose the agent applied on each sweep. Prose drifts. Now each rule is a pure function with a passing test, a failing test and a mutation check, and the docs say plainly that the pacing intervals were picked by hand, not measured.

  • The pacer holds back lower-priority cards so only one decision reaches me at a time.
  • Almost all commits in the repository were written by agents under my review.
  • Automated tests and mutation checks break one pacing rule at a time to prove the suite catches drift.
  • The app runs on a deliberately tiny set of runtime dependencies.