Skip to main content

dify-docs-translation-test

Judge a zh or ja page as the owner of that edition would: a fresh agent reads it beside its English source against the language's formatting guide and the glossary, marks the sentences that fall short, and returns a ship verdict on the work of this round. Runs in the pipeline's Check stage after the editor test.

Source facts

Repository
langgenius/dify-docs
Last source activity
October 28, 2026 at 03:43
Detected SKILL.md language
English
Stars
180
Forks
399

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
name
dify-docs-translation-test
description
Judge a zh or ja page as the owner of that edition would: a fresh agent reads it beside its English source against the language's formatting guide and the glossary, marks the sentences that fall short, and returns a ship verdict on the work of this round. Runs in the pipeline's Check stage after the editor test.
# Translation Test The CJK linter measures mechanics. This test measures the translation: whether it says what the English says, in the language a native reader would write. It only measures anything if the judge has not seen the translating: a fresh sub-agent gets the translated page, its English source, the language's guide, and the glossary, and nothing else. ## Procedure 1. Agree the dispatch with the owner before sending anything: one judge per language per round is the usual shape, reading every page that round touched in that language; a page translated whole can have its own judge when the round is large. 2. Dispatch the fresh sub-agent(s) (`subagent_type: general-purpose`) with the template below, filling `{LANG}` (Chinese or Japanese), `{GUIDE}` (`tools/translate/formatting-zh.md` or `tools/translate/formatting-ja.md`), `{GLOSSARY}` (`writing-guides/glossary.md`), `{SOURCE}` (the English page, absolute path), `{DRAFT}` (the translated page, absolute path), and `{CHANGED}` ("the whole page", or the sections and paragraphs this round changed). When one judge reads several pages, the page block (files 3 and 4 and the changed line) and the reply repeat per page, under the page's path. Put nothing else in the prompt. 3. Relay the report unedited. Then act on it: the verdict judges this round's work; "Ship with light edits" means fix the marked sentences, "Needs heavy edits" means retranslate the marked units whole from the English, "Needs retranslation" means a fresh translator takes the page from the English file on disk. A revised page goes to a new judge. Marks outside the change go to the owner as a list with a proposal for each. ## Dispatch prompt ```text You are the editor of the {LANG} edition of a documentation set, reading a translated page the way its owner would before deciding whether it can ship. Read exactly these files, in this order: 1. {GUIDE}: the whole guide. Its translation-quality sections describe the failures you are looking for; the rest of the guide is context for them, and its formatting rules are checked elsewhere. 2. {GLOSSARY}: the terms and UI labels, with their {LANG} forms. Then, for each page under review: 3. {SOURCE}: the English page. It is what the translation has to say. 4. {DRAFT}: the {LANG} page under review. This round changed: {CHANGED}. Read the whole page regardless; a sentence is judged in the context the page gives it. Read nothing else and run no other command. Mark every sentence that falls short, quoting it, with one tag each. A sentence, block (a list, table, code block, callout, or image), or section of the English page with no counterpart in the translation is marked too, quoting the English: - meaning: says something the English does not, or drops something it does (a condition, a restriction, a qualifier such as "only", a whole sentence, block, or section) - structure: carries the English clause structure or sentence order where a native writer would not - register: a word in the wrong register for its context - term: a general term whose {LANG} form is not the glossary's, or a UI label that does not correspond to the label the English page names - tell: a translationese pattern the guide lists - rule: breaks a translation rule the guide states and the tags above do not name (a standard phrase, a katakana convention, an app-type name); name the section Then give one verdict on this round's work: Ship with light edits / Needs heavy edits / Needs retranslation, with one sentence on what decided it. Marks outside the change do not move the verdict; list them separately. Do not check formatting, links, or facts beyond what the English page says. Do not rewrite. Judge only whether a native reader would take this page for one written in {LANG}. Reply with exactly this, once per page, under the page's path: - **Marked, in this change**: [quoted sentence, or the missing English unit — tag — five words on why; for rule, the section], one per line, or "none" - **Marked, elsewhere on the page**: same form, or "none" - **Best passage**: [the section that reads as written, in a phrase] - **Verdict**: [one of the three, and the deciding sentence] ```
View on GitHub