diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..38c166c
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,54 @@
+# Notes for Claude
+
+Things that are true of this repository and cost somebody a round trip to
+find out. Not a style guide — `CONTRIBUTING.md` is that.
+
+## Translations
+
+Nine languages ship alongside English: German, Spanish, French, Dutch,
+Portuguese (Brazil), Russian, Ukrainian, Simplified Chinese and Japanese, in
+`web/src/locales/`. A missing key renders its English source rather than
+failing, so an untranslated string is invisible until somebody reading that
+language finds it.
+
+**Any change that adds or alters a user-visible string adds work in all nine
+catalogues.** Say so explicitly when reporting the change — how many keys, and
+the fallback count before and after — and say so just as explicitly when a
+change adds none, so it is never left to be inferred.
+
+### The catalogue key for a plural is the `other` form
+
+`plural()` looks the entry up by `forms.other`, so a call site written as
+
+```ts
+plural(n, { one: "Deleted {n} contact", other: "Deleted {n} contacts" })
+```
+
+is keyed on **`"Deleted {n} contacts"`**. Keying the catalogue on the `one`
+form type-checks, builds, passes every test, and silently falls back to English
+in all nine languages. Nothing errors. The only signal is the fallback count
+going up, so read it:
+
+```sh
+npm run i18n:check                    # literals wrapped, and catalogue health
+node scripts/i18n-catalog-check.mjs   # per-language: translated / used / falling back
+```
+
+Compare the "falling back to English" number against `main` before and after.
+It should not rise. Do not read the percentage instead — adding keys moves the
+denominator, so it can hold steady while new strings go untranslated.
+
+Plural forms are per language, from `Intl.PluralRules`: `one`/`other` for most,
+`one`/`few`/`many`/`other` for Russian and Ukrainian, `other` alone for Japanese
+and Chinese. Supplying a form a language does not draw is inventing a
+distinction, not being thorough.
+
+## Verifying UI work
+
+Store tests do not exercise the component. At least one bug in this repo's
+history — a shift-click range measured inside a `setState` updater, which React
+runs after the anchor ref has already moved — passed every store assertion and
+failed the moment the built app was driven. If a change is visible on screen,
+run it: `npm run dev:mock` (mock Stalwart, credentials printed on start), then
+drive the real thing. Add a component test for what you find; there are
+examples in `web/src/views/*/__tests__/`.