/ˈlæŋɡwə ʒɛm/

A progressive language-immersion drip for Claude Code. It preprompts, forces and monitors how much of each reply comes back in a target language, and ramps that share up on evidence you're keeping up, not on a calendar.

Standalone: one Python file, stdlib only, no pip installs. A blend drip, not a hard language switch.

How it works

Four layers, wired as three Claude Code hooks plus a statusline segment:

• Preprompt (SessionStart hook). languageme hook injects the drip into every new session: blend ~N% of your prose into the target, keep the rest in your usual reply language. It states N as a concrete per-paragraph quantum (a model cannot self-calibrate a raw percentage) and tells you to judge each reply against the target, not against your own earlier replies. N lives in state. The target is confined to chat prose, never code, commits, filenames, tool arguments, or any clause where a wrong word breaks something.
• Nudge (UserPromptSubmit hook). languageme nudge closes the loop the one-shot preprompt cannot. Before each reply it injects one live line from the previous turn's measured blend: last reply 80% vs ~16% target, well over, one short phrase per paragraph. This re-asserts the target every turn (immune to context growth) and corrects both overshoot and full dropout, the two ways an open-loop drip drifts.
• Monitor (Stop hook). languageme monitor reads the finished transcript, measures the real target-language share you produced (char-weighted, per-token classifier that reads the intra-sentence blend, no ML dependency), tracks whether you asked for help or wrote the language yourself, and moves N per the ramp policy. The ramp reads the session-cumulative blend and gates on a band: too much target counts as noncompliance, not mastery, so overshoot never ramps N up.

A statusline segment rides along so you always see where the dial is. Reading sv 12%↑ ~11:

• sv : target language
• 12% : current blend target (what the preprompt aims for)
• ↑ : the gauge just moved up on the last turn (↓ down, nothing = steady)
• ~11 : the real share of target language measured on the last turn (~ = roughly). No ~N yet means no turn has been measured.

install wires it non-destructively: if you already have a statusline it wraps it and appends the segment, if you don't it adds one. It never edits or replaces your statusline.

The blend only ever grows when the real measured share holds the target and you are not asking for help.

Quick start

./languageme init sv          # seed Swedish at 10%
./languageme install          # wire hooks + statusline (edits ~/.claude/settings.json)
./languageme status           # target, blend %, recent measurements, ramp state

The preprompt loads at SessionStart, so the drip starts on your next Claude Code session. To activate it in the current one without leaving, run /clear (it re-fires SessionStart). Until the first turn is measured the statusline shows the target with no ~N, and the gauge stays put; the ~N and any ramp appear once a turn has run under the drip.

Put it on your PATH if you like: ln -s "$PWD/languageme" ~/bin/languageme. The hooks call it by absolute path, so a symlink is optional.

Switch target any time: languageme lang sv|es|lv. The gauge and ramp follow the new language.

Ramp policies (languageme ramp <mode>)

• mastery (default). Ramps one step after 3 distinct sessions whose cumulative blend lands in the target band (not too little, not too much) with zero help asked. Each session counts once, however many turns it ran. Writing the language yourself adds a bonus step. Overshoot or asking for help resets the streak.
• calendar. One step per monitored session until the goal. Hands-off, blind to whether you're keeping up.
• manual. Never auto-moves. You bump it: languageme bump +5.

Natural-language overrides always win, both directions, and they are language-aware. Say "too much swedish" (or the target's own words, "för mycket svenska", "demasiado español", "trop de français") in chat and the next monitor pass steps the blend down; say "more swedish" / "mer svenska" / "más español" and it steps up. Asking to switch back ("in english", "en français") or to gloss a word ("what does X mean") counts as needing help and holds the ramp. Ordinary coding talk ("translate the config", "the query got slower") does not.

Help modes (languageme help <mode>)

• auto (default). Glosses new words as word (translation) on first use, fades as they recur.
• on. Every target phrase glossed. zero ambiguity, slower learning.
• off. Raw, infer from context.

Add languageme phonetic on to layer a short pronunciation respelling on the tricky words (the å/ä/ö vowels, sj/tj/rs clusters, silent letters), e.g. sköldpadda (turtle, "HYULD-pad-da"). Plain learner respelling, not IPA, only where the spelling hides the sound. Off by default; ignored when help is off.

Pause and kill switch

Two levels, so you can always shut it off:

• Persistent toggle. languageme off pauses the drip (no preprompt, no measuring; the statusline shows sv off). languageme on resumes it. The state persists across sessions.
• Hard kill (env). export LANGUAGEME_OFF=1 fully disables it for that shell or session, overriding state no matter what. The hook injects nothing, the monitor does nothing, and the statusline falls back to pure passthrough (your original line, no segment). The guaranteed off switch.

Commands

init <lang>                 seed a target language (sv, lv, de, es, it, ...)
status                      full state readout
statusline [--wrap]         compact segment; --wrap runs the wrapped statusline too
hook                        [SessionStart] print the preprompt
monitor [--session P]       [Stop] measure a turn, ramp the %
        [--full] [--quiet]  re-measure whole transcript / silence the report
set pct|goal|floor|step <n> tune the numbers
bump [+n]                   nudge the blend now
goal <n>                    ceiling to ramp toward
help on|off|auto            inline-gloss policy
phonetic on|off             pronunciation cues on tricky words (å/ä/ö...)
ramp mastery|calendar|manual
lang [<code>]               show / switch target
off | on                    pause / resume the drip (persists)
install | uninstall         wire / unwire hooks + statusline
update                      self-update (git pull the tool's checkout)

Measurable targets: sv, es, lv. Others get the drip but no % yet.
Hard kill for a single shell/session (overrides everything): LANGUAGEME_OFF=1

State

Lives in $LANGUAGEME_HOME (default ~/.languageme/state.json). The code is portable; the state is the only thing that mutates. install backs up your settings to settings.json.languageme.bak before its first edit.

Notes

• Measurable targets: sv, es, lv (plus fr/en as discriminators). Each ships a stopword lexicon and its distinctive characters in the LEXICONS / LANG_CHARS registries; the target is scored against all the others. Adding a language is those two entries, nothing else. The drip and preprompt work for any language out of the box, only the measurement half is lexicon-bound (an unmeasured target just holds the % at 0).
• install appends its hook entries and wraps (never replaces) your statusline; it leaves every existing hook and statusline untouched. uninstall removes only its own and restores the original statusline.
• Scales flat. State is bounded (history and session cursors are capped), each Stop reads only the new transcript tail via a byte seek (not the whole file), and the read-modify-write is flock-serialized so parallel Claude Code sessions can't clobber each other's measurements.

License

MIT. See LICENSE. Use it, fork it, ship it.