Post-editing MT that's technically correct but stylistically wrong takes just as long as translating from scratch.

The post-editing promise didn't work out the way it was supposed to.

The expectation was reasonable: AI translates 80%, a human corrects 20%. The reality is more frustrating than that. When a generic MT engine produces output that's grammatically accurate but completely wrong for the client's brand — formal where the brief said casual, "Bildschirm" where the style guide specified "Monitor," "Nutzer" where the client uses "Kunde" throughout — fixing it takes longer than starting over. The AI saved no time; it just moved the work.

The problem isn't the translator. It's that generic MT has no knowledge of this client. It doesn't know that "Dashboard" should stay untranslated in German, that the register is casual du not formal Sie, or that the client's style guide uses ten specific phrases that must appear verbatim. That context doesn't exist anywhere in the pipeline — so the output is missing it, and the post-editor has to supply it by hand.

From the community:

The fix isn't a faster MT engine. It's a pipeline that loads the client context before generating anything.

Your client's glossary is enforced before the first word is generated.

When Flixu translates a project, it loads the client's Glossary as a payload constraint before the language model receives the source text. The approved term isn't a suggestion the model can override when the context gets ambiguous — it's specified in the input. The model builds the surrounding grammar around the fixed term from the start.

In practice: "Monitor" appears as "Monitor." "Kunde" appears as "Kunde." The term you would have had to manually correct fifteen times across a 5,000-word document isn't there to correct — because the constraint prevented it before generation.

When the client updates their terminology, the change applies immediately to every subsequent translation. No email thread, no version of the style guide that the previous translator had but you don't.

→ How glossary constraints work: Glossary Enforcement

Your client's brand voice is in the configuration, not in your mental model.

A client's brand voice guide is a PDF. You read it, internalize it, try to apply it across a 10,000-word project, and somewhere around hour four, the formality level drifts. Not because you're not good at your job — because holding a complete stylistic model in memory while also making translation decisions is genuinely hard.

The Brand Voice Manager stores the client's tone configuration in the workspace. Formality level, phrasing preferences, stylistic constraints — defined once, applied automatically to every translation request for that client. When you open a project for them, the brand voice is already active. You're not applying it from memory; you're reviewing whether the output met a standard that the system was enforcing.

For freelancers managing multiple clients with different brand voices, this means the configuration travels with the client profile rather than with your mental cache.

→ Brand voice configuration: Brand Voice Manager

Structural elements — variables, tags, placeholders — survive translation intact.

One of the most common frustrations in post-editing software files is the variable problem. {username} becomes {benutzername} in the translated output. <strong> becomes <kräftig>. The code breaks. Someone has to find all the broken placeholders and fix them — and the error doesn't always surface until the file is deployed.

Flixu's parser extracts only the translatable text before the language model sees anything. JSON keys, HTML tags, Markdown syntax, and interpolation variables are preserved in the source structure and re-injected into the output file. The translated file is structurally identical to the source — the model only translated the human-readable text.

For freelancers working with software strings, documentation, or any structured file format, this means the output is review-ready, not repair-ready.

→ File format support: Document Translation

How Flixu fits into a freelance workflow.

What's different about using context-aware translation as a freelancer.

The difference between Flixu and running a file through DeepL or Google Translate isn't the translation quality in isolation. DeepL produces fluent output. The gap is what the model knows when it generates that output.

Generic MT generates text without knowing the client. It doesn't know their approved terms, their preferred register, their brand voice, or what you approved last month. The output is statistically likely — which for a specific client with specific requirements means wrong in specific, predictable ways.

Flixu generates text with the client context already loaded. The model receives the source text with the glossary constraints, brand voice parameters, and Translation Memory references already assembled. The errors that would have required post-editing — wrong terminology, wrong register, missing style — were prevented before generation.

For a freelancer, this shifts the review from "fix what's wrong" to "verify what's right." The time difference is real — and it compounds across every project for that client as the Translation Memory deepens.

Supported file formats.

For each format, Flixu's parser preserves the structural elements — keys, tags, variables, code blocks, Markdown syntax — and extracts only the translatable text for translation. The output file is structurally identical to the source.

→ Full file format documentation: Document Translation