Experiment
Ambient language immersion
Claude weaves a few words from a language you are learning into the replies you were going to read anyway. A word arrives with its translation the first few times. Once you have seen it enough, the translation stops coming.
Korean, for example
Both tests pass, so the 계획 (gyehoek) holds. after a dozen sightings, the same sentence reads: so the 계획 holdsHangeul earns that middle step, where the word stands on its own but keeps its romanisation. A language written in the Latin alphabet has nothing to romanise, so the Spanish version of the same sentence goes from the plan (el plan) straight to el plan.
How it works
A scheduler picks each day's words and decides how each one is written, from a file of per-word exposure counts. That division is the whole design: a model asked to run its own curriculum drifts, quietly and in the direction of whatever the conversation is about, while a hundred lines of Python does not.
Asking what a word means is the other half. The question counts as a miss, the word drops back to showing its translation, and the schedule absorbs that without you tracking anything.
What you get
- No language comes preloaded, so the first run asks for yours and writes the word list.
- Your progress lives in one file on your own machine, outside the skill.
- Nothing is sent anywhere, and the scheduler uses the standard library alone.
- Deleting the directory removes the whole thing.
Install
Then ask Claude Code to run the immersion skill. The first session asks which language you want, writes a starter word list for it, and begins. The bundled README covers making it always-on.
sha256 eefeacb63035461c177fcbe25321d9e95dcf0522e870cb42fdd813d08d973769 · checksum file