I save interesting posts on X the way other people scroll: as a break. An assistant then triages the saves, so I asked it for a short list of what I should read myself. Its first answer was to cap the list at three, because that matched how often I had actually read anything it flagged. The reasoning was tidy, and it was wrong in a useful way.
The better argument for a long list is that my taste and the filter’s taste are not the same. A tight filter applies its own judgement and hides its misses, so I never see what it left out. A generous filter keeps those items on the page, where my reading can overrule it. The first evidence arrived the same afternoon. The triage had marked a Gwern essay and two Noah Smith essays as not worth adopting, which was true for my work and irrelevant to whether I would enjoy them. A work verdict had been quietly standing in for a reading verdict.
Recall only works with two guards. The list has to be ordered, so I can read from the top and stop anywhere and the tail costs nothing. Unread items also have to leave on their own. I left my read-later app this year with more than two thousand articles still unread in its queue. Making saving effortless is how that happens, and a recall-first list without expiry rebuilds the same backlog with better intentions.
Two smaller choices mattered more than I expected. The list shows when each item was added, never when it will leave, because a visible deadline turns a menu into a debt, and the debt is the fear of missing out the list was meant to quiet. It also keeps work reading out. An essay I would use to prepare an answer at work is not a break, however good it is. Good writing about my own field can still belong when I would read it for its own sake, and those pieces sit in their own short section, so finishing the top few feels like being caught up.
Ordering by my taste raised one more problem. A list ranked by what I already read drifts toward what I already think. So each section keeps one strong piece near the top that argues against a view I hold, marked so I can choose between a challenge and a rest. An agreeing piece earns its place by adding evidence or a mechanism, not by restating my view well.
None of this needed a reading app. The problem was never where I read. It was choosing what to read, and a ranked list that forgets is enough for that.