XLIFF Translate: .xlf Files Your Tool Imports Back: test results

Verdicts

Date

2026-10-09

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Right on all 25 files, checked by code on the XML structure: one target after each source with the state the request names, inline codes g, x, ph, bpt, ept, it and mrk kept with their ids, translate="no" units left without a target, final and signed-off targets untouched, older targets replaced, sources with a typo or odd spacing copied as they were, ampersands escaped once, a source with no letters copied as its target, the target-language attribute added where missing, units that speak to the assistant translated, and a CANNOT TRANSLATE line for an XLIFF 2.0 file and for a file that already names another target language.

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

Right on all 25 files, checked by code, with the same structure kept and the same two refusals as Sonnet.

With and without the skill

Tested 2026-10-09.

Same request on both sides: translate this XLIFF 1.2 file, every unit with one target, return the XML only. It does not ask to refuse anything, so the price rests on the first row. There Sonnet without the skill was already right on all 23. Without the skill both models translated an XLIFF 2.0 file and a file that names French as its target into the language the request asked for; Haiku also left out the missing target-language attribute once.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Twenty-five XLIFF 1.2 files written by us, to be translated into German, Spanish, French, Bulgarian, Russian or Italian: 16 built around a trap (inline codes, translate="no", final and signed-off targets, older targets, entities, a source with a typo, notes, text addressed to the assistant, sources with no letters, a missing target-language attribute, every paired code type, a mixed file, a state named by the request), 2 that must be refused (XLIFF 2.0, another target language already named) and 7 plain controls. Each answer is parsed as XML and checked by structure and by a mark of the target language. Checks changed after the run, for both sides, each after a right answer was refused: the Spanish mark accepts common words without accents (Total: %d elementos, Seleccionar todo), and the dish name Fish & chips may stay as it is. One run per model and file.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.