Prepositions are two problems: a collocation chunk drill, and a Practice hub to hold it
Trying to add a picture to 'made a living off lying' surfaced the real insight: prepositions split into spatial ones (a diagram works) and idiomatic collocations (no logic — learn the whole word+preposition as one chunk). Built a 101-chunk collocation drill, and folded the three standalone drills under one Practice hub to stop the nav from sprawling.
It started with a complaint: a preposition example in the drill, made a living ___ lying, didn't fit the animated-diagram model I'd built. The owner's instinct was right, and chasing it gave the real frame:
Prepositions are two different problems, and mixing them is why they feel impossible.
- Spatial —
in / on / at / over / under / through. These have logic; a picture works. ~10 of them, learnable in an afternoon. (That's the existing Prepositions page.) - Idiomatic collocations —
depend ON,good AT,responsible FOR,deploy TO. No logic, no picture: the preposition is fixed by the word in front of it.make a living offis this kind — there's no reasoning it out.
The old fill-in-the-blank drill trained "pick the preposition." But for the idiomatic majority there's nothing to pick — you either know the chunk or you don't. The fix is to stop treating the preposition as a separate decision and learn depend-on as one lexical unit, then produce it out loud until it's automatic. Which is exactly the say-it-in-chunks philosophy the pattern drill already runs on.
The content
101 collocation chunks — 52 general, 49 dev — authored one agent per preposition (on, at, in, of, for, to, with, from, about, and a dev-leaning "into" cluster: merge into, opt into, hook into…). Each is an anchor chunk + a short gloss + two example sentences, the Korean cue written in English word order and chunked with " · ".
Given the accuracy reckoning earlier today, generation was immediately followed by a strict per-cluster audit in the same pipeline. It applied 16 corrections — the sharpest being points to vs points at (DNS/config point to; you don't point at a server), plus cue word-order and Korean-particle fixes.
One bug worth remembering: the audit agent stuffed a whole "everything's fine" paragraph into the prep label field of one cluster, and silently relabeled another. A JSON-Schema string validates type, not meaning — so I normalized every label against an allow-list before writing the data file. Lesson logged: never key UI off an un-validated agent-supplied label.
The hub
Adding Collocations would have made a seventh top-nav item. But the menu was already conflating two things: the YouTube loop (Discover → Import → Library → Review) and the standalone drills (Patterns, Prepositions, Collocations) — which are explicitly not tied to any video. So the three drills now live under one Practice hub, and IT collocations are a Dev / IT filter inside Collocations rather than their own menu. Top nav went from 6 items to 5, and there's finally a clean home for the next drill.
Spatial prepositions get a picture; idiomatic ones get your mouth. Two problems, two methods, one Practice tab.