localization-release-copilot
Shadow-mode copilot for localization / translation release operations. Given a source draft and its target-language version(s), it locks terminology, machine-checks numbers and proper nouns, preserves source attributions, flags structure loss, and produces a reviewable diff preview — without rewriting the translation. Designed for editorial long-form ZH↔EN localization; SaaS locale-file features (.po / placeholders / plurals) are flagged as pending until real samples exist.
Localization Release Copilot
For a localization / translation release flow, take the source draft and its target-language version(s) and act as a shadow reviewer that catches release-breaking drift before publish: terminology drift, number mismatches, dropped proper nouns, lost source attributions, and structure loss. Output only a reviewable diff and a preview-ready checklist — never rewrite the translation itself.
When to use
- You have a finished source (ZH article, paper, docs page, release note) and a human or model translation into another language (EN / others), and you want a pre-publish consistency pass.
- You publish the same content in multiple languages (WeChat ZH → LinkedIn EN / Substack EN; docs ZH → EN; in-product copy → locale) and want to confirm the target did not silently drop a number, a proper noun, or a source citation.
- You maintain a glossary / term base and need every release to respect locked terms (product names, framework names, units, branded metrics).
- You want a reviewable PR-style preview of what changed between source and target so a human reviewer can approve in one pass.
- You suspect "the English version reads fine but lost the data" — this skill is built to catch exactly that class of bug.
Do not use
- A single-language copyedit / proofread of one draft → use
claim-to-source-auditoror a dedicated compliance-review skill if one is installed. - Consistency across different materials on the same topic (PR vs pitch vs deck) → use
cross-material-consistency-auditor. - Topic selection / portfolio evaluation → use
editorial-topic-portfolio. - Machine translation generation → this skill reviews, it does not translate.
- Directly modifying or rewriting the target text.
Input
source:
path:
language: zh | en | ...
targets:
- path:
language:
label: LinkedIn | Substack | docs-EN | locale-EN
- path:
glossary: # optional: locked terms the target must respect
- source_term:
target_term:
mandatory: true | false
preserve_attribution: true # default true
review_only: true # never rewrite; output a diff + preview
Workflow
Step 1: [Deterministic] Load source + targets
- Read the source and every target.
- Detect language, length, section count, and status (draft / final).
- If a glossary is provided, load it as the locked-term table.
Step 2: [Deterministic] Extract & normalize
For each file extract:
- Numbers with units, normalized to a canonical magnitude (locale-aware):
- ZH
亿→ ×1e8,万→ ×1e4; ENB/billion→ ×1e9,M/million→ ×1e6,K/thousand→ ×1e3. 8.3亿≈830M,117亿≈€11.7B,1.5亿≈€150Mare equivalent, not mismatches.
- ZH
- Proper nouns / entities (capitalized tokens, quoted terms, org names, person names, product names).
- Source attributions (
(Source, Year),Source (Year), inline citations). - Glossary terms (if a glossary is supplied).
- Section-level structural promises (headings, promised lists).
Output 01-extracted-claims.json.
Step 3: [LLM] Build the diff matrix
Compare source → each target item by item; for each inconsistency record:
- Dimension:
terminology|number|proper_noun|attribution|structure_loss|locale_format. - Source wording vs Target wording.
- Difference type: term not honored, number dropped/changed, entity missing, citation lost, section omitted, correct locale adaptation misread as error.
- Severity:
- P0: a number or legal/factual claim in the source is missing or contradicted in the target.
- P1: a source attribution (org + year) is dropped or changed; a mandatory glossary term is not honored; a key proper noun is missing.
- P2: non-mandatory term drift; section depth mismatch that does not change the claim; platform-specific condensation.
- Recommended unified wording (cite the source).
- Whether re-check is needed after the target is revised.
Consistent items are kept as positive records but not listed line by line.
Step 4: [LLM] Produce a reviewable preview
Render a PR-style preview the human reviewer can approve in one pass:
- A short summary: source → target(s), what passed, what needs attention.
- A line-level diff for each P0/P1: source sentence → target sentence → suggested fix (do not auto-apply).
- A glossary-compliance block: which locked terms were honored / violated.
- A "ready to publish?" gate: blocked if any P0 remains, needs review if any P1 remains, clear otherwise.
Step 5: [Deterministic] Human confirmation
Pause after the preview. Confirm:
- P0/P1 disposition.
- Whether the target should be revised (this skill never revises it).
- Whether the glossary needs an update.
After confirmation, the human or the corresponding Writer/Editor edits the target. This skill is review-only.
Hard Rules
- Review-only. Never modify the source or the target. Output a diff + preview; hands-off after delivery.
- Locale-aware by default.
8.3亿=830M,117亿=€11.7Bare correct localizations — never flag them as mismatches. A checker that is not locale-aware will produce false positives and destroy trust. - Attribution is part of the content. If the source cites
(Synergy Research, 2024), the target must retain both the org and the year. Dropping either is a P1. - Glossary terms are mandatory when marked. A locked product/framework name must appear in the target exactly as specified.
- P0 blocks publish. A missing or contradicted number/factual claim in the target is release-breaking.
- Source is authoritative. When source and target disagree on a fact, the source wins for arbitration; the target is the one to fix.
- No fabricated wording. Recommended fixes must reference an existing source string, never invent.
Failure Handling
| Scenario | Action |
|---|---|
| Only the source provided, no target | Stop; the task requires at least one target |
| Target unreadable | Mark it and check only the remaining targets |
| Source contains an obvious error | Flag the error; do not propagate it into the target recommendation |
| No glossary supplied | Run with terminology inferred from the source; note "glossary recommended" |
| Source and target are different documents, not a translation pair | Stop and warn: this skill needs a true source→target pair, not two related-but-different docs |
| All checks pass | Report "no localization drift found" and clear |
| P0 found in an already-published target | Flag with extra severity and recommend a correction |
Output Format
<run-dir>/
├── 01-extracted-claims.json
├── 02-diff-matrix.md
├── 03-preview.md
└── 04-glossary-compliance.json
Verification
- Every target compared to the source on numbers, proper nouns, attribution, glossary, structure.
- Every P0/P1 has a recommended fix that cites the source.
- Locale-equivalent numbers (亿/M, 万/K) treated as matches, not errors.
- Glossary-compliance block present for every mandatory term.
- Preview contains no source/target modifications.
- Human confirmation documented before the task is marked complete.
Pitfalls
- Treating correct locale adaptation (
8.3亿→830M) as a mismatch — the #1 way to make this skill useless. - Honoring the target's "reads fine" and skipping the number check — drift hides in fluent text.
- Using two related-but-different documents (e.g. a ZH long article and an EN short commentary on the same theory) as a translation pair — they are not, and the check will produce noise.
- Assuming the model translation is the source of truth — the source draft is authoritative, the target is what gets fixed.
- Forgetting that LinkedIn/Substack condensations are legitimate; only flag structure loss that drops a claim, not a section.
中文摘要(Chinese Summary)
本 Skill 是本地化 / 翻译发布运营的影子助手:给它源稿和其一/多个目标语言版本,它锁定术语、机检数字与专有名词、保留来源标注、标记结构损失,并产出可审阅的差异预览——不重写译文本身。
当前范围(Scope): 聚焦长文 editorial 的 ZH↔EN 本地化子集(源稿→目标稿)。SaaS locale 文件特性(.po diff、占位符、复数、翻译记忆)标注为"待真实样本补齐",本版不依赖这些文件即可运行。
关键约束(双语要点 / Bilingual key points):
- 只审阅不改写 Review-only:绝不修改源稿或译文,只输出差异矩阵与预览;交付后 hands-off。
- locale-aware 是生命线 Locale-aware:
8.3亿=830M、117亿=€11.7B是正确本地化,误报这类等价会被当成 bug 直接毁掉可信度。 - 来源标注是内容一部分 Attribution matters:源稿
(Synergy Research, 2024)在目标稿必须保留机构+年份,掉落任一项即 P1。 - 术语表强制 Glossary mandatory:标了 mandatory 的锁定词(产品名/框架名/单位)必须原样出现在目标稿。
- P0 阻断发布 P0 blocks:目标稿缺失或歪曲源稿的数字/事实主张,即为发布阻断级。
- 源稿为权威 Source authoritative:源稿与目标事实冲突时以源稿仲裁,要改的是目标稿。
何时该用它(场景引导 / When to reach for it): 你有成型的中文源稿 + 英文(或其他语言)译文,想在发布前确认"英文版没悄悄丢数字 / 专有名词 / 来源";你维护术语表需要每次发布都遵守;你想要一个可一键审阅的 PR 式预览。它专门抓"英文读着通顺但数据丢了"这类 bug。
- Downloads
- Security audit
- Pass
- Last updated
- 1w ago
- Current version
- v1.0.1
- License
- MIT-0
