Five Target Languages, One Claude Haiku 4.5 Pipeline

01 · One Doc Set, Five Target Languages and a Single Release Train

Once a product team commits to five languages, documentation stops being a writing chore and becomes a production line. Onboarding guides, release notes, API references and troubleshooting pages all have to land in each language at the same time as the English source. The familiar failure mode is delay: English ships first, translations trickle in over the following weeks, support absorbs the gap, and writers spend their days reviewing drafts instead of writing new material.

The fix is not more translators. It is treating machine translation as a real production stage with defined inputs, quality gates and a budget line. When the source is stable, a model pass produces drafts in all five languages quickly, and people spend their time where judgement counts: terminology, tone, screenshots, and anything with legal or compliance weight. The operating gains appear in predictable places:

• › All five languages advance together, so release notes can be published in every market on the same day.
• › Reviewers open a near-final draft instead of a blank page, which changes the review conversation entirely.
• › New pages inherit the existing pipeline instead of waiting for a translator to become free.
• › Throughput becomes measurable per language, so you can see where the process actually stalls.

02 · Choosing the Model for Bulk Drafts and the Pages That Carry Risk

Not every page deserves the same treatment. A long FAQ or a changelog is low-risk, high-volume material where speed matters more than nuance. A licensing page, a security overview, or a migration guide full of deprecation language is the opposite: small volume, high consequence. Splitting the corpus by risk before you touch a model keeps both cost and error rates under control.

A workable split for many teams is to use a fast, inexpensive model for the bulk pass across all five languages, and to reserve a larger, more careful model for the small set of pages where wording can create commitments. In our own planning that maps to Claude Haiku 4.5 for bulk drafts and Claude Opus 5.5 for the risk-tier pages. Whatever you pick, write the split down as a rule so the decision is not renegotiated every sprint.

• › Bulk tier: reference pages, FAQs and release notes, where a stylistic imperfection is cheap to fix.
• › Risk tier: security and privacy statements, entitlements, limits, and anything a customer could quote back to you.
• › Human tier: executive posts, launch claims, and pages a named reviewer must sign off before publication.

03 · Glossary Enforcement and Cache Reads That Pay Off

Terminology is where five-language programs quietly fail. The same feature ends up with three names across locales, translators invent synonyms for user interface labels, and support answers questions about a product that does not exist in the customer's language. Build the glossary before the first draft: approved product names, interface strings exactly as they appear on screen, prohibited translations, and a short note on register for each target language.

Feed that glossary to the model as part of every request and check the output against it mechanically before a human reads a single line. The second, underrated lever is shared context. Style guide, glossary and a handful of reference pages form a long prefix that is identical across hundreds of calls. Because cache writes and cache reads are priced separately on real token usage, that repeated prefix becomes much cheaper on every call after the first. At team scale this matters more than it sounds: a documentation cycle is sustained, parallel calling, not occasional traffic.

Keep the unit economics honest. Track source characters per language, draft length after translation, and the share of calls that hit the cached prefix. Teams that only count request volume discover the real cost after the invoice rather than before it.

04 · Human Review Gates That Catch Register Drift

Model drafts fail in characteristic ways: false friends, register that is too formal or too casual for the product voice, units and dates left in the source convention, and sentences that are correct but read like a translation. A review gate is worth building only if it catches those specific classes. Random sampling catches almost none of them.

Structure review in three passes. First, a terminology check against the glossary, done automatically before any human time is spent. Second, a native-speaker read of a fixed sample per language per release, drawn from the pages that changed most. Third, a targeted review of the risk tier, where a person signs off every line. Feed reviewer notes back into the glossary so the same error cannot return next cycle.

• › Ask reviewers to mark register problems separately from accuracy problems; the two have different fixes.
• › Track reviewer edits per hundred lines by language. The language with the highest edit rate needs a better prompt, not more reviewers.
• › Keep a list of strings that must never be retranslated, such as product names and command names.

05 · Metering, Budgets and a Two-Week Rollout Sequence

Roll out in two-week cycles instead of one big-bang launch. Week one: pick a corpus slice, build the glossary, define the tiers, and run a pilot in two languages end to end. Week two: extend to all five, wire in the review gates, and measure. The pilot exists to reveal workflow problems, not to prove that a model can translate.

Budget with numbers from the pilot rather than guesses. Since every call is metered on real tokens, the pilot gives you a defensible cost per page per language, and you can extrapolate the full corpus with confidence. Keep a separate allowance for rework: most teams find that first-pass drafts need one editing round and the risk tier needs two. Watch the cache hit rate on your shared prefix as well, since it moves the cost curve more than model choice usually does.

Sustained workloads are exactly what pay-as-you-go relays are built for. Quanta runs pipelines like this on one balance across a wide range of models, with cache writes and cache reads priced separately so the economics stay legible as volume grows. Open an account and run the two-language pilot against your own corpus before committing the full five-language cycle.