Principles

You can see why the card in front of you appeared: the thread you pulled, the field you chose to stay inside, the page you asked to circle, or “drifting” when nothing was chosen. There is a line on the card saying which.

What this rules out: a recommender. The usual way to improve a feed is to learn what holds someone's attention and serve more of it without saying so. Drift keeps a light interest model, but it only affects which broad areas turn up when you drift at random, it is shown on the card, and it can be switched off. The working rule is that if the reason cannot be stated in about four words on the card, the feature does not get built.

Nothing advances on its own. No autoplay, no timers, no card sliding partway into view.

What this rules out: a deep queue. Drift fetches at most one card ahead. That is a deliberate limit rather than a technical one: a large buffer is what makes a feed feel like there is always more waiting, and once twenty cards are ready, stopping starts to feel wasteful. One ahead is enough for the next card to appear instantly.

A session has a beginning (you pick where to start), a middle (the trail), and an end (a map of where you went).

What this rules out: putting the payoff at the next swipe. The trail map is the most interesting part of the app and it is only reachable by stopping. Moving it earlier would very likely increase time spent, and would remove the reason it works.

There is a count of your stops in the corner. After about 25, Drift mentions once that you have been going a while. You can dismiss it and it will not ask again.

What this rules out: most of the standard engagement toolkit.

• No streaks.
• No badges or unread counts.
• No push notifications. Drift does not contact you at all.
• No daily goal and no time-spent target.

The counter is deliberately the whole of it. Anything stronger tends to become pressure of a different kind rather than less pressure.

Everything you read in Drift was written by people, for an encyclopedia or a museum catalogue. Drift cuts it into cards and works out the threads between them. It does not write the facts.

What this rules out: generated text on a card. Where a language model is used at all it is optional, runs locally, and is limited to labelling and ordering material a person already wrote. If it is unavailable the app behaves the same way, because no card depends on it. More on sources.

Mostly because they are easy to lose gradually. Nearly every change that would make Drift more habit-forming is small and locally reasonable, and the ones above are the specific things I would otherwise talk myself into.

If you think Drift has broken one of these, let me know.