Crowdin

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build the translation workflow I actually need instead of Crowdin — and know what makes this ALMOST.

**Translation memory.** Every sentence you have already paid to translate, reused automatically the next time it appears. That is what makes the second language cheaper than the first, and it is a real asset that lives in this kind of tool. Losing access to it is the migration risk, not the software.

STACK
- Node 20+ with Fastify, server-rendered HTML
- SQLite through better-sqlite3, WAL mode, with FTS5
- Caddy in front

THE DATA MODEL
- strings: id, key, source_text, context, file, line, screenshot_path, is_plural, max_length, first_seen_at, removed_at
- translations: id, string_id, locale, text, status, translator_id, reviewed_by, updated_at — draft, translated, reviewed, approved
- memory: id, source_text, locale, text, source_project, used_count — **the asset, and it is exportable in a standard format or it is not yours**
- glossary: id, term, locale, translation, note, is_forbidden — the product names that must never be translated and the ones that must be translated one way
- Everything append-only enough that a translation's history is readable

GETTING STRINGS IN AND OUT
- Extract from the repository on every merge: the key, the source text, **the file and line**, and any comment the developer left
- Push translations back as a pull request rather than committing directly. That review step is what stops a broken placeholder reaching production
- Support the string formats you actually use — one per framework — with a fixture and a test each. This is the dull part and it is where the bugs are
- A string removed from the code is marked removed, never deleted. It comes back, and its translation with it

WHAT A TRANSLATOR ACTUALLY NEEDS, AND MOST TOOLS UNDERSERVE
- **Context.** The same English word is a noun and a verb, and 'Post' is either. A screenshot of where the string appears, or at minimum the developer's comment, is the difference between a good translation and a guess
- The maximum length, when the string sits in a button that cannot grow
- **Placeholders shown as placeholders**, and a check that refuses a translation which dropped one or renamed it. That single validation prevents most translation-caused crashes
- Plurals as the target language defines them — several languages have more than two forms, and a system with only singular and plural is broken in them from the start

MACHINE TRANSLATION, HONESTLY
- Useful as a first draft and labelled as one, in the data as well as the interface
- **Never published without a person clearing it**, and never for legal or safety text at all
- Translate with the surrounding strings as context rather than line by line, or pronouns and formality fall apart

THE MEMORY
- Every approved translation goes in. A new string matching one already there is offered with its match percentage
- **Exportable in a standard interchange format.** That is what makes the asset yours, and it is the first thing to check about any tool in this category, including this one

WHAT MATTERS MOST
Context for the translator and placeholder validation. The first decides the quality of what you get back; the second decides whether it crashes the application.