Translations as Build Artifacts, Not Source Files

articleailocalizationi18nnextjsreactdeveloper-toolsagent-workflows

Treating translations as generated artifacts maps to joelclaw's preference for source-of-truth code plus agent-authored PRs instead of parallel config drift.

Translation files are one of those chores that pretend to be architecture. General Translation is making a clean bet: write the UI where it lives, then let GT extract, translate, and deliver localized output as build/runtime infrastructure instead of making humans babysit JSON dictionaries.

The clever bit is source-of-truth inversion. In i18n without translation files, they describe generated translation files as output artifacts, not files developers author and maintain. Locadex pushes that into agent territory: connect GitHub, let an AI agent understand project structure, internationalize code, and open pull requests with translations.

That’s not “AI writes copy” fluff. It’s CI/CD for language drift.

For joelclaw, the useful pattern is treating multilingual content as a pipeline: code, docs, and MDX are source; localized files are cache; pull requests and previews are review points. Same shape as any decent agent workflow: detect change, produce patch, preview, verify, ship. The risk is obvious too: if the translation system owns context and delivery, we need receipts, previews, and rollback. Otherwise we traded JSON file misery for spooky action at a distance.

Key Ideas

  • General Translation treats application code and components as the source of truth for localization instead of requiring hand-maintained per-locale translation files.
  • GT is open source and includes packages for Next.js, React, React Native, Sanity, a CLI, compiler integrations, and an MCP server.
  • Locadex is the agent-shaped piece: it connects to a codebase, understands project structure, and opens localization pull requests.
  • The homepage describes routing, PR-aware deployment, preview workflows, edge delivery through a translation CDN, and live translation for user-generated content.
  • The pattern to steal for joelclaw is source → generated artifacts → preview → review → deploy, with explicit receipts so generated content doesn’t become invisible infrastructure bullshit.