1 files changed, 61 insertions, 0 deletions
diff --git a/TODO-ui-languages.md b/TODO-ui-languages.md
new file mode 100644
index 0000000..f12c32b
--- /dev/null
+++ b/TODO-ui-languages.md
@@ -0,0 +1,61 @@
+# Translate the interface
+
+## What exists today
+
+`web/src/languages.js` holds 16 **typing** languages: the vocabulary the
+generated-word modes (Words, Timed, Zen, Practice) draw from. The picker in
+the top-right rail sets `settings.language`, and `wordGenerator.js` reads it.
+
+That is the same thing MonkeyType and 10FastFingers mean by "language" - what
+you type, not what the buttons say. The interface itself is English only.
+
+## What this task is
+
+Making the interface follow that picker, or a second picker beside it.
+
+## Why it is not a small change
+
+Every user-facing string is currently a literal inside the markup that renders
+it. There is no catalogue, no lookup, and no plural or interpolation handling.
+The work is:
+
+1. **Extract.** Pull every string out of `web/src/**` into a catalogue keyed by
+   id. Roughly 250 strings across screens, tooltips, empty states and errors.
+   Tooltips are the bulk of it and the easiest to miss, because they live in
+   `data-tooltip` attributes rather than in text nodes.
+2. **A lookup with interpolation and plurals.** `"{n} racing"` and
+   `"{n} players racing now"` already need a plural rule, and Polish, Czech and
+   Romanian have more plural forms than English does. Use `Intl.PluralRules`
+   rather than an `n === 1` check.
+3. **Decide whether the two languages are one setting.** Typing Spanish does
+   not mean wanting a Spanish interface, and vice versa. Two pickers is more
+   honest; one is fewer controls. This is a product decision, not a technical
+   one.
+4. **Translate.** 16 locales. This is the part that cannot be done from inside
+   this repo with any confidence - see below.
+5. **Layout.** German and Finnish compound words are long. The corner rails,
+   the mode picker and the end screen's stat labels are all sized to English
+   today; several will need to wrap or shrink.
+
+## The honest constraint
+
+Machine-translating 16 locales and shipping them as finished would put text in
+front of users that nobody who speaks those languages has read. The word lists
+have the same issue but a much smaller blast radius: a slightly odd word in a
+typing drill is a curiosity, whereas a mistranslated button is a bug the user
+cannot diagnose.
+
+Suggested approach:
+
+- Build the catalogue and lookup, with English as the only complete locale.
+- Add one or two further locales fully, to prove the plumbing against real
+  translated text of realistic length.
+- Ship the rest only as native speakers review them, with the picker listing
+  only locales that are actually complete.
+
+## Related
+
+- `web/src/languages.js` - typing vocabulary, and the note there about the
+  lists being curated rather than corpus-derived.
+- The corpus question is separate and also open: real frequency lists are
+  mostly CC BY-SA, which does not sit cleanly in an MIT repo.