1,956 example sentences the verb pack never had — generated in parallel, merged at build
The base-verb pack showed a Korean cue and an English pattern but no sentence using it. Generated one natural example + Korean translation for all 1,956 entries with a 103-agent workflow, and merged them as committed side-data so the hand-editable source stays clean.
"예문이 1도없네" — not a single example. The founder opened a Base-verbs card, saw be into [topic] with a Korean gloss, and nothing showing the phrase actually spoken. He was right, and I checked before I answered this time: VerbItem had no example field, the generated phrasal-verbs.ts had zero "example" keys, and the source default_verb_v3.md never carried examples. The pack was a pattern list, not a sentence bank.
Generating 1,956 of them
One agent per verb group (103 groups, run in parallel), each asked for a single natural example sentence plus a Korean translation for every entry in its group — real standup/chat/daily English, brackets filled with concrete content, no textbook tone. The output holds up:
be into [topic]→ "My roommate is really into rock climbing these days." / "내 룸메이트가 요즘 암벽 등반에 완전 빠져 있어."*have ... under control→ "Don't worry, I have the release under control." / "걱정 마, 릴리스는 내가 통제하고 있어."*do the trick→ "Just restart the router — that usually does the trick." / "그냥 공유기 재시작해 봐, 보통 그러면 해결돼."*
103 agents, ~1.7M tokens, every group answered, no gaps: 1,956 / 1,956 merged.
Where the examples live
The interesting decision was not to write the sentences back into default_verb_v3.md. That file is the human-editable source of truth, and the next build-verb-pack.py run would either fight the AI text or the parser would choke on it. Instead the pairs are committed as keyed side-data — scripts/data/verb-examples.json, keyed "<groupId>#<index>" — and the build script merges them onto each item after parsing the markdown. So the source stays clean and re-generatable, and the expensive AI pass survives every rebuild. It's the same pattern the pack already used for particle-classifications.json.
VerbItem gained example? / exampleKo?, and the drill reveal now shows the sentence (and its Korean) under the model — on both the Base-verbs screen and the Today session. Shipped on feat/mobile-app (067fb49), build 17 to TestFlight.