One developer, nine apps, and Claude at every step

Claude Code builds every app

Each repository has a CLAUDE.md / AGENTS.md that records its conventions, its decisions and the reasons behind them. Claude Code reads that file and works inside it: product notes, features, tests, Supabase Edge Functions, release scripts.

• New apps start from a shared Expo template that comes with i18n, design tokens, onboarding, a paywall, RevenueCat and server-side quotas already wired in. Stirred and Remate were both created from it.
• Store metadata, screenshot pipelines and fastlane releases are written and kept up to date the same way, for both the App Store and Google Play.
• Four App Store releases shipped in the first ten days of October 2026.

Claude Opus localizes our apps

Our i18n-sync script checks every locale against English. It reports missing keys, leftover keys and dropped {{tokens}}, and it flags any locale that is just an untranslated copy of English. With --fill it sends the gaps to the Claude API and writes the translations back. Stirred ships in 31 languages this way.

• Tokens must survive translation. A dropped {{count}} breaks a sentence at runtime, and only a reader of that language would notice.
• Plural forms: Claude adds the extra CLDR forms each language needs (_few, _many for Slavic languages, all six for Arabic).
• Strings are translated for UI, not literally. They have to fit the button they sit on.

node scripts/i18n-sync.mjs --check            # CI / preflight
node scripts/i18n-sync.mjs --fill --lang de,fr  # translate the gaps with Claude

Claude handles the studio's operations work

An internal tool built on the Anthropic SDK scores freelance listings and open-source bounties, then drafts a proposal or a technical assessment for each one. A person reviews and sends everything; nothing is posted automatically.

Milo: what a word means in its sentence

A reader taps a word in a quote they saved and gets the dictionary form, the part of speech, the sense used in that sentence, and one line of explanation in their own language. "kederinden" comes back as "keder", and "bank" next to a river means the shore.

• Hard cases: Turkish inflected and archaic words, English words explained to a Turkish reader, and five other book languages.
• The integration is built on Claude Haiku 5.5 with structured output. On our 30-word eval it got 29 right; on the one it missed, it said it didn't know instead of inventing a meaning. It costs about $0.0001 per lookup.
• No personal data reaches the model: a request carries only the word, the sentence and the answer language. Each answer is computed once and cached.

Stirred: shelf scan and reasons behind each pick

Shelf scan turns a photo of a home bar into catalog IDs using an enum-constrained schema. We plan to test Claude's vision on the same shelf-photo eval set, and to have Claude write the one-line reason behind each night's "for you tonight" suggestion in the user's language.

HairOn: from onboarding answers to a style brief

HairOn's onboarding asks for face shape, goals, interests and timing. Claude would turn those answers into a structured brief that picks and ranks the looks we generate, in each of HairOn's 31 languages. Image generation itself stays on image models.