i18n and copy

Who is this for

A developer or copy editor who needs to change a text on the landing page, publish or hide the Uzbek version, or add a language. It also explains to a translator why a “simple” wording change can fail the build.

TL;DR

• Every visible string lives in src/i18n/dictionaries/{en,ru,uz}.ts. en.ts defines the Dictionary type; ru and uz must match it key for key. No hard-coded strings in components.
• npm test rejects forbidden words, any digits other than 12, 72 and 3D, wrong apostrophes in Uzbek, a breakable space before a dash, and the wrong form of address.
• Each language is written as an original, not a translation. Russian addresses the organiser as «вы» on the page and the guest as «ты» inside product mock-ups; Uzbek is «siz» everywhere.
• Uzbek is published (uzPublished: true) but never the default: no browser language leads to /uz.
• A longer text can fail npm run audit: the H1 may take at most 3 lines at 1280 and 1920 and 4 lines at 360 and 768, and header elements must have identical geometry in all three languages.
• The Uzbek dictionary has not been proofread by a native speaker yet. See Known gaps.

Where texts live

File	Role
src/i18n/dictionaries/en.ts	English. Its shape is the Dictionary type
src/i18n/dictionaries/ru.ts	Russian, typed Dictionary
src/i18n/dictionaries/uz.ts	Uzbek in Latin script, typed Dictionary
src/i18n/index.ts	export type Dictionary = typeof en and getDictionary(locale)
src/i18n/dictionaries.test.ts	The rules below
src/content/site.ts	Things that are not words and read the same in every language: studio name and link, the partner agency credit and its on/off switch, the year, the uzPublished flag

Top-level keys

Key	Covers
meta	page title and description (also used for Open Graph)
nav, header, localeSwitch, footer	chrome
hero	H1, subtitle, CTA, scene caption, stepper labels, one sentence per scene step
scene	product mock-ups: kiosk screens, bot messages, the group chat in “Why”
showcase, stickers	showcase texts, people descriptions, the 12 sticker labels
how, why, organizer, data, setup, faq	page sections
form	labels, placeholders, error messages, consent line, “sent” screen

scene is the only key with a different voice. Everything outside it talks to the event organiser. scene is the product talking to a guest.

How to change a text

1. Find the key in en.ts. Change the string in all three files. Each language is an original: rewrite it so it reads naturally, do not translate word for word.
2. If you add a key, add it to en.ts first. TypeScript then reports the missing key in ru.ts and uz.ts.
3. If you add an array item, add it to all three. The type does not check array lengths; the test does.
4. Put a no-break space (written   in the source) before every em dash. See the dash rule below.
5. Pass the string to components through props. A client component receives only its slice: Pick<Dictionary, "form">, not the whole dictionary.
6. Run the checks:

npx tsc --noEmit
npm test
npm run shots -- --only=360,1280
npm run audit

shots and audit need a running server. See Local setup — landing page.

Never write a string literal in a component. The only texts drawn by code are things that are not copy: the clock time inside Telegram mock-ups and the 3-2-1 countdown on the kiosk.

Tone and form of address

Area	Russian	Uzbek	English
The whole page: hero, sections, FAQ, form, “sent” screen	«вы», to the organiser; the guest is spoken of in the third person	«siz»	direct; “you” is the organiser
Product mock-ups (every key under scene): the bot and kiosk talking to a guest	«ты», as the real product does	«siz»	—

Further tone rules, from the repository’s CLAUDE.md:

• Short sentences, active voice, concrete statements.
• The page says “we”. It never names the company. The partner agency appears only in the footer credit, which comes from SITE.partner and is switched off with partner.enabled.
• The reader is the organiser who gets a turnkey service. Do not write that they “run” or “manage” the station.
• Mock-ups are in the page’s language, including the pack title.
• The sticker style is described in words (“chunky cartoon”, «объёмные мультяшные», «hajmli multfilm»). Its internal name never appears.

What may and may not be promised

May be said	Must not be said or named
QR without an app, entry through Telegram	prices
guest consent before the photo	any figure except “12 stickers”, “a couple of minutes” and “72 hours”
choosing an outfit on the tablet	reviews and client logos
12 chunky cartoon stickers with the guest’s face	other messengers
the pack in Telegram in a couple of minutes	printing stickers
the station is free immediately, it does not wait for generation	“works without internet”
one pack per Telegram account	“100 % likeness”
photos deleted automatically within 72 hours	names of AI models and providers
monitoring and technical support during the event	a backup or fallback AI provider
the station in the colours and with the logos of the event	branded outfits or props
setup for the event with a rehearsal	brands of past events

What the dictionary test enforces

npm test runs these against all three dictionaries.

Same shape, nothing empty

• ru and uz have exactly the same key paths and array lengths as en.
• No string is empty or whitespace only.

Forbidden words

The whole dictionary is serialised to JSON, lower-cased and searched for substrings. Two consequences:

• Key names count too, not only values.
• Substrings match. A harmless word that contains a forbidden one fails. For example, any English word containing print or magic, or any word containing the letters gpt in sequence.

Category	Blocked substrings
Another messenger	one messenger brand name
Printing stickers	печать стикер, печатаем, print, chop etish
Working without internet	без интернета, offline, internetsiz
Guaranteed likeness	100%, 100 %
AI models and providers	three provider and model names
Brands of past events	two brand names
A backup provider	резервн, запасн, fallback, backup, zaxira
Hype	революц, revolution, inqilob, новый уровень, next level, all-in-one, магия, magic, sehr
The style’s internal name	stylized
The partner agency	its short name

The brand names are deliberately not repeated here. The exact list is the forbidden array in the test file.

Allowed digits

Every run of digits, optionally followed by one letter, is checked. Only three tokens pass:

Token	Meaning
12	stickers in a pack
72	hours until photos are deleted
3D	allowed as a token

One exception: in the keys form.phone.placeholder and form.errors.phoneInvalid a sample phone number starting with +998 is removed before the check, because it shows the format to type and promises nothing.

So “24/7”, “5 minutes”, “2026”, “1st” and “No. 1” all fail. Write numbers as words if you must mention them, and check first whether you may promise them at all.

Uzbek apostrophes

Letter	Required code point	Rejected
oʻ, gʻ (also capital)	U+02BB ʻ (modifier letter turned comma)	U+02BC after o or g
tutuq belgisi, as in maʼlumot	U+02BC ʼ (modifier letter apostrophe)	U+02BB after any letter other than o or g
—	—	ASCII ', typographic ‘ and ’, backtick, anywhere in uz

Most keyboards and translation tools produce the rejected characters. Paste the correct ones from an existing string. The fonts load only the latin and cyrillic subsets, and both code points are in latin; this is also why the monospace font is IBM Plex Mono, which has ʻ.

The dash rule

A dash must never start a line. The test rejects a normal space followed by an em dash ( —) in any string. Write a no-break space (U+00A0) before the dash. The dictionaries use the escape   so the character is visible in review:

title: "AI Sticker Pack — personal stickers for every guest at your event",

Form of address

The test each part keeps one form of address checks whole words, case-insensitively:

Dictionary part	Must not contain
ru, every key not under scene	ты, тебя, тебе, тобой, твой and its forms
ru, every key under scene	вы, вас, вам, вами, ваш and its forms
uz, every key	sen, seni, senga, sening, sendan, senda

English has no such check.

Routing rules in the same file

The test file also pins the routing functions:

• Accept-Language resolution for eleven sample headers;
• six Uzbek headers that must never resolve to uz;
• publishedLocales and languageAlternates for both values of the flag;
• that the site’s own setting is currently true.

Locale routing

Defined in src/proxy.ts and src/i18n/config.ts. In Next.js 16 the file is called proxy.ts; it plays the role earlier versions called middleware.

flowchart TD
    A["Request path"] --> B{"Starts with /ru, /uz or /en?"}
    B -- yes --> C["Serve the page"]
    B -- no --> D{"NEXT_LOCALE cookie is a published locale?"}
    D -- yes --> E["Redirect to the cookie's locale"]
    D -- no --> F{"Accept-Language, by quality"}
    F -- "first match is a CIS language" --> G["Redirect to /ru"]
    F -- "first match is English" --> H["Redirect to /en"]
    F -- "only other languages" --> H
    F -- "missing, empty or a wildcard" --> G

Rule	Detail
Locales	ru, uz, en, in the order the switch shows them
Default	ru, used when the browser says nothing useful
CIS languages mapped to /ru	ru, uz, uk, be, kk, ky, tg, tk, az, hy — by primary subtag, so uz-Latn-UZ and kk resolve alike
Quality values	honoured: en;q=0.3, kk;q=0.9 gives /ru
Uzbek browser	/ru. The Uzbek page is chosen by the visitor, never detected
Cookie	NEXT_LOCALE, written by the language switch, one year, SameSite=Lax. It records only the chosen language
Paths left alone	anything under /_next, /api, and any path containing a dot (static files)
Other unprefixed paths	redirected to /<locale><path>
hreflang	one entry per published locale, plus x-default pointing at /en
Language switch	swaps the first path segment, keeps the hash and the scroll position, cross-fades with a view transition

The uzPublished flag

One boolean in src/content/site.ts:

locales: {
  uzPublished: true,
},

Effect on	true (current)	false
/uz by direct link	opens	opens
Search indexing of /uz	indexable	robots: noindex, nofollow in the page metadata
Language switch on /ru and /en	RU / UZ / EN	RU / EN
Language switch on /uz itself	RU / UZ / EN	RU / UZ / EN — the draft still shows itself lit, for the proofreader
hreflang alternates	ru, uz, en, x-default	ru, en, x-default
sitemap.xml	lists /uz	omits /uz
Cookie NEXT_LOCALE=uz on /	redirects to /uz	ignored; Accept-Language decides
Browser language Uzbek on /	/ru	/ru

How the checks follow the flag:

• npm run audit reads the flag from site.ts and expects the running build to behave accordingly. npm run audit -- --only=uz runs just the language and theme groups, for trying the other value on a second build.
• npm test checks the routing functions for both values. It also asserts that the committed value is true (test the site's own setting is what the defaults follow — and it is published). If you set the flag to false, update that assertion in the same commit, or npm test fails.
• The flag is read at build time. Changing it needs a new build and deploy.

Layout constraints from the longest language

Text length is a layout input here. Two different languages are “the longest”, depending on the element.

Element	Longest language	Constraint	Enforced by
Overall copy	Uzbek	it defines the hero text column (44 %), the header slots and the CTA width	convention
H1	Russian	the top of the H1 size clamp, 2.35rem, was chosen for the Russian headline	audit
H1 line count	all three	≤ 4 lines at 360 and 768; ≤ 3 lines at 1280 and 1920	audit
H1 line count through the font swap	all three	identical in the fallback font and in Unbounded	audit
Hero subtitle	all three	≤ 4 lines at 360 and 768; ≤ 2 lines at 1280 and 1920	audit
Header CTA and hero CTA	all three	identical x, width and height in every language at 1280 and 1920; text not clipped	audit
Navigation links	published locales	centre drifts at most 1 px between languages	audit
Kiosk title in the hero tablet	all three	one line at 768, 1280 and 1920	audit
Hero CTA on a phone	all three	visible without scrolling at 360×780	audit
Sentence under the stepper	per locale	every step has a distinct sentence	audit (on /ru)

The H1 font size is clamp(1.55rem, 1rem + 2vw, 2.35rem). If a new headline does not fit, shorten the headline. Do not raise a line limit to make a text pass.

Adding a new locale

Caution

Inferred from code, not yet done in practice. The site has shipped with exactly three locales since its first version; nobody has added a fourth. Treat this as a checklist of every place the code names a locale, and expect to find one more.

1. Config. In src/i18n/config.ts: add the code to locales; extend the hard-coded comparison in isLocale; add an entry to OG_LOCALE; decide how Accept-Language should treat the language (today the function returns only ru or en) and whether the new locale needs its own published flag like uzPublished.
2. Dictionary. Create src/i18n/dictionaries/<code>.ts typed as Dictionary and register it in src/i18n/index.ts.
3. Switch label. Add localeSwitch.<code> to every dictionary; the switch renders dict.localeSwitch[code].
4. Switch component. In src/components/site/LocaleSwitch.tsx extend the ORDER array and the path regular expression /^\/(ru|uz|en)/.
5. Fonts. Check that Unbounded, Onest and IBM Plex Mono cover the script with the loaded subsets (latin, cyrillic). Adding a subset changes font weight on the first screen; re-measure.
6. Tests. In dictionaries.test.ts add the dictionary to ALL, add the language’s forbidden words and its form-of-address rule, and update the expected arrays in the publishedLocales and languageAlternates tests.
7. Scripts. Update the hard-coded locale lists: LOCALES in scripts/audit.mjs and scripts/shots.mjs (and the file-name pattern (ru|uz|en) in shots.mjs), the loop in scripts/js-weight.mjs, the default runs in scripts/lighthouse.mjs, the list in scripts/scroll-film.mjs. The audit’s expected switch contents (ru,uz,en) change too.
8. Layout. Run npm run shots and npm run audit. A new language may be the new longest one; the header geometry and H1 limits apply to it from the first run.
9. Leads server. The server accepts locale values ru, uz, en and stores anything else as en (LOCALES in the product repository’s bot/services/leads.py). Add the code there, or requests from the new page are recorded as English.
10. Things that follow automatically because they iterate locales or publishedLocales(): the redirect in proxy.ts, generateStaticParams, the sitemap, the hreflang alternates, the Open Graph image route.
11. Product language. The real bot and kiosk speak Russian only. A landing page in a new language shows mock-ups in that language; the product has to follow before the first event in it.

Proofreading status of Uzbek

• The Uzbek dictionary was written as a draft. Its header comment and one comment in the form section (К вычитке, “to be proofread”) say a native speaker’s review is pending.
• The page is nevertheless published and indexable (uzPublished: true).
• Treat every Uzbek string as unreviewed. When the review happens, re-check the three typographic rules above, because reviewers’ tools tend to replace ʻ and ʼ with ordinary apostrophes.

The gap is tracked on Known gaps.

Where to go next

• Landing quality tooling — the audit checks that constrain text length.
• i18n across the system — languages in the bot, the kiosk and the landing page together.
• Showcase packs — outfit names, people descriptions and sticker labels are dictionary keys too.
• Known gaps — the pending Uzbek proofread and other open items.
• Client FAQ — what sales may promise, in line with the “may be said” list.

---

Last verified: 2026-11-13 against ai-sticker-pack@b715418 and ai-stickers-landing@e350ff5.