Localise

Localise

English is the canonical source. The target-language deliverable is derived and review-flagged. This skill produces a target-language version of a finished English deliverable, a target-language companion, or a review of your own target-language draft. It does not author from scratch in the target language.

Supported languages: Romanian (ro, fully curated), French (fr), Spanish (es), Russian (ru). French, Spanish, and Russian are working-draft glossaries refined at use. Adding a further language is a new glossary file.

When to use

Trigger for: localising a finished English brief; producing a target-language companion to an English deliverable; reviewing a target-language draft you wrote. Do not trigger for English voice work (use ane-voice) or instrument design (use instrument-designer).

Modes (detected from input)

1. Localise a finished doc then you provide an English deliverable (paste or path) and get a target-language version plus a QA note.
2. English plus companion then Vi or you provide the finished English deliverable and get a target-language companion alongside it.
3. Review my draft then you provide your target-language draft (plus the English source if available) and corrections are walked one by one, never a silent rewrite.
4. Review an image-based deck export then you provide a slide deck or e-learning module exported to Word (Articulate Storyline, PowerPoint), where the translatable text is rasterized inside the slide images and the body holds only slide titles. You get anchored Word comments on a copy plus a branded slide-keyed corrections sheet. Read references/deck-export-mode.md before running it.

Mode 4 triggers when a review target is a .docx whose text lives in images rather than in the body. Confirm it rather than assuming: python scripts/extract_deck.py extract DECK.docx --out-dir slides/ prints the verdict with the evidence behind it, and routes you back to mode 3 when the text is live. The linguistic core (glossary, controlled terminology, back-translation, safeguarding) is identical to mode 3; only the input and output machinery differ.

Required inputs

1. The English source (or, for mode 3, the target-language draft plus the English source). Required.
2. Target language: Romanian, French, Spanish, or Russian (Romanian is the most curated; the others are working drafts).
3. Audience and member-association context (shapes register and cultural adaptation).
4. Restrictive-context flag (triggers the safeguarding pass on sensitive terms). Mandatory for Russian-language output (see the legal-review banner in the Russian glossary).

Method

1. Read the target-language glossary mel_wiki/wiki/glossaries/srhr-terminology-<lang>.md, substituting the target-language code (ro, fr, es, ru). For Russian, read the legal-review banner at the top first.
2. Translate and adapt. Apply controlled terms; preserve BLUF, plain language, collaborative voice, numbered steps; adapt examples to the member-association context.
3. Terminology-compliance check. Confirm every glossary English term in the source uses the controlled target-language rendering.
4. Back-translation QA. Back-translate key passages to English; diff against the source; flag meaning drift.
5. Safeguarding pass. List every sensitive-flagged rendering used; in a restrictive context, surface the term choice for confirmation and route it to safeguarding-reviewer (mandatory for the legally-exposed Russian terms: LGBTI+, gender identity, sexual orientation, CSE, reproductive rights).
6. Output. IPPF-branded Word via ane_package.reporting.word_export with the document language set via the lang parameter (ro/fr/es/ru); plus the QA note.

Steps 1 to 5 are identical in mode 4. Step 6 is not, because the corrections belong in the source deck rather than in a Word file: follow references/deck-export-mode.md for the extract, comment and sheet steps.

Output structure

1. The target-language deliverable.
2. Terminology-compliance note.
3. Back-translation QA note (meaning-drift flags).
4. Sensitive-term list and the choices made.
5. Native-review stamp ("native review recommended" / "validated by author").
6. Data gaps.

Mode 4 replaces item 1 with two artefacts: the commented copy of the export (the original is never touched) and the branded slide-keyed corrections sheet, which is the actionable one. Items 2 to 6 are unchanged.

Citation requirements

ITC (2017) ITC Guidelines for Translating and Adapting Tests 2nd ed; Brislin (1970) back-translation (canonical); WHO process of translation and adaptation. Controlled terms cite their glossary source.

Writing rules

Follow CLAUDE.md house style for the English QA notes. The target-language prose follows the equivalent Tier 1 register in that language. Default register is Tier 1 working brief unless Ane names publication.

Limitations

Produces a draft for native review where the author cannot validate the output. Does not replace native-speaker sign-off for languages the author does not read. Romanian is the most curated glossary; French, Spanish, and Russian are working drafts whose unsourced and competing-rendering terms are refined at first use. Russian carries restrictive-context legal-exposure flags — never distribute Russian-language SRHR material inside Russia without legal review and safeguarding-reviewer sign-off.

Edit-preservation protocol

In mode 3, and any time Ane references an existing file by path, read the file first, edit scope-bounded via the Edit tool, preserve out-of-scope content byte-identical, and return the EDIT-PRESERVATION DELIVERY summary.

Apply mel_wiki/wiki/concepts/edit-preservation-protocol.md when target file exists.