Voice → Notion → AI Classifier
Soundcore Work × Notion × Claude automation pipeline. Voice recordings are auto-classified into Diary, Ideas, and To-dos, with monthly summaries generated on top. Idempotent design via a Processed Log DB prevents duplicates.
A voice recorder that writes straight into Notion sounds ideal until you use it for two weeks and the workspace is a landfill of untitled recordings — diary entries, product ideas and to-dos all in one undifferentiated pile.
The fix was to stop treating the recorder as a destination and treat it as an input stream. Soundcore Work feeds Make.com, Claude classifies each transcript into diary / idea / to-do, and each class lands in its own Notion database. A monthly rollup then summarises what actually came up.
The design point that mattered was idempotency. The first version had no dedupe, so every scenario run reprocessed the same pages and entries multiplied. The rewrite added a Processed Log database: check before doing anything, let only unseen items through, and write the log at the end of each branch. Without that step an automation is not a time-saver — it is a source of incidents.
It runs on free tiers plus one AI call per recording, which is a few hundred yen a month. No server, no scheduler, nothing to keep alive. The monthly summary turned out to be the most valuable output: what you actually thought about this month, from data rather than from a flattering memory.
Stack
