Use when creating, reviewing, or maintaining public-domain book translation projects and EPUB pipeline templates in this repository. Follow multilingual, language-pair-specific, rights-aware workflow rules.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Use when creating, reviewing, or maintaining public-domain book translation projects and EPUB pipeline templates in this repository. Follow multilingual, language-pair-specific, rights-aware workflow rules.
Use this skill when an AI agent is asked to create a book project, add or update a language-pair template, revise prompts, review policy documents, or build EPUB output in this repository.
当 AI agent 被要求创建书籍工程、新增或更新语言方向模板、修改 prompt、审查政策文档或构建 EPUB 时，使用本 skill。
For new or migrated translation projects, treat translation_units/ as the only authoritative source-target store. Lock the full-source proper-noun discovery manifest, candidate decisions, entity_id register, occurrence ledger, terminology, and translation contract before parallel translation. When the user does not choose a proper-noun display mode, policy 3 is allowed only with selection_source=default; it never waives discovery or review. Translators write {{pn:entity_id}} templates in worker-owned chapter patches, and the single CAS merger renders names from locked CSV data. chapters/translated and chapters/final are deterministic identical-target projections and must not be edited independently.
For long-book parallel work, read references/adaptive_parallel_orchestration.md, run the weighted planner, and use Scheme B: disjoint translation producers, independent chapter audit consumers, and one coordinator/merger. Never infer capacity from page count alone. Unknown or unauthorized client capability means zero spawned workers. GPT-family clients may spawn at most four workers and non-GPT clients at most eight; start with no more than two and scale only after the pilot satisfies the recorded structural, proper-noun, semantic, rework, and conflict thresholds. A reviewer must not be that chapter's translation owner. Different chapters use chapter-scoped CAS and audit invalidation; same-chapter stale patches still fail.
When the active client exposes safe subagent creation, its coordinator must populate the capability file from live tool/model metadata, run npm run translation:orchestration:plan, and create no more workers than the resulting spawned_worker_count. Keep the returned task/agent IDs in run evidence and enforce the planned producer/auditor roles. Re-run the planner after the pilot, rate-limit changes, budget changes, or quality regressions. If the client cannot create workers, continue serially; do not emulate parallel writes with uncontrolled processes.
活动客户端具备安全子代理创建能力时，coordinator 必须依据实时 tool/model metadata 填写能力文件，运行 npm run translation:orchestration:plan，且实际创建数不得超过 spawned_worker_count。返回的 task/agent ID 必须进入执行证据，并严格执行规划中的 producer/auditor 角色。pilot 完成、限流变化、预算变化或质量退化后必须重跑规划器。客户端不能创建 worker 时应串行继续，不得用不受控进程模拟并行写入。
For bilingual editions, one complete source short paragraph must be followed immediately by its complete target paragraph. A unit may split one long natural paragraph only at complete-sentence boundaries; it must never cross two source natural paragraphs. Sentence-by-sentence subtitles, viewport/page chunking, or several source units followed by a target-language batch are release blockers. Structural PASS and EPUBCheck never substitute for hash-bound semantic omission/addition/neighbor-boundary audits.
Ordinary translation and builds use full-book static DOM/text/order/navigation checks and EPUBCheck only. A real-reader smoke test is attempted only for the final release candidate. If no supported reader is installed, record SKIPPED_UNAVAILABLE, allow release, and disclose the unverified boundary to the user; never skip the static gates.
Before doing any repository task, read the relevant template/ files. The template is the source of truth; do not rely on memory or prior runs.
执行任何仓库任务前，必须读取相关 template/ 文件。模板是规则源，不得依赖记忆或历史执行经验。
Read the repository README in the user's preferred language when available.
优先读取用户偏好语言对应的 README。
For EPUB production, cover, book-info/frontmatter, assets, quality gates, random review, or release work, read the applicable common policy files before editing or building:
Covers use the concise producer line LifeBook 书坊 译制. Personal contributor names belong in book-info.xhtml and metadata only when the book-info policy requires them.
Treat large-chapter shrinkage and note loss as pipeline defects, not as model-style problems. The durable fix is a deterministic coverage gate, not a one-off prompt reminder.
大章后段缩水和注释丢失应视为流水线缺陷，不是单纯模型风格问题。长期解决必须依靠确定性覆盖门禁，而不是一次性 prompt 提醒。
check_translation_coverage.py is the shared gate. It checks matching chapter filenames across chapters/src/, chapters/translated/, and chapters/final/. It fails on low paragraph-block coverage, low non-space character coverage for long chapters, note reference/definition loss, unresolved target note references, and lost table/image/formula units.
check_translation_coverage.py 是共享门禁。它按同名章节对比 chapters/src/、chapters/translated/ 和 chapters/final/。若段落块覆盖率过低、长章非空白字符覆盖率过低、注释引用/定义丢失、译文注号没有定义，或表格/图片/公式单元丢失，必须失败。
preflight:template and check:chapter-controls must include this gate. If an Agent writes a PASS chapter control while the coverage gate fails, the PASS is invalid and the chapter must return to translation or reconstruction before any EPUB build or release.
preflight:template 和 check:chapter-controls 必须包含该门禁。若 Agent 写出 PASS 的章节 control，但覆盖门禁失败，该 PASS 无效；章节必须回到翻译或重组后才能进入 EPUB 构建或发布。
For very long chapters, split translation work into stable paragraph or small-block units and keep notes with the relevant unit or a dedicated note unit. The final merge must not proceed until every unit and note definition has a target counterpart.
对超长章节，应按稳定段落或小块单元翻译，并让注释跟随相关单元或独立注释单元。所有单元和注释定义都有对应译文前，不得合并为终稿。
Language Requirements / 语言要求
Important human-facing files must include the local language for the intended contributors.
面向人的重要文件必须包含目标贡献者能读懂的本地语言。
English may be included in parallel as a bridge language, but English-only important instructions are not acceptable unless English is the template's contributor language.
英文可以作为桥接语言并列出现，但除非英语就是该模板贡献者语言，否则重要说明不能只写英文。
For English-to-Japanese, use Japanese plus optional English. For German-to-Traditional-Chinese, use Traditional Chinese plus optional English. For French-to-English, English is acceptable.
Shared repository-level documentation should include Chinese when project-owner review is expected.
需要项目发起者审阅的共享仓库级文档应包含中文。
Hard Stops / 必须停止的情况
Rights or public-domain status is unclear.
版权或公版状态不清楚。
The only available source is a modern copyrighted translation or unclear commercial EPUB.
唯一来源是现代受版权保护译本或来源不明的商业 EPUB。
The current working directory is still under template/ while book-specific output is about to be written.
当前目录仍在 template/ 下，却准备写入具体书籍产物。
A pretranslation report, chapter gate, EPUB validation, stratified random spot-check, closure validation, independent review, or PASS release gate fails.
预翻译报告、章节门禁、EPUB 校验、分层随机抽检、闭环校验、独立评审或 PASS 发布门禁失败。
Publication lint reports hard errors such as mojibake, legacy print page-number tables, abnormal spacing, or target-language punctuation violations.
出版文本 lint 报出硬错误，例如乱码、旧纸书页码目录、异常空格或目标语言标点违规。
Long printed-TOC chapter titles are still being used as EPUB navigation labels without title design.
旧纸书目录式长章节标题仍被直接用作 EPUB 目录题名，尚未完成标题设计。
Source-language names or parenthetical originals appear in translated chapter titles, subtitles, or navigation labels where the target-language title should stand alone.
译文章节标题、副标题或目录题名中出现原文人名或原文括注，而目标语言标题本应独立呈现。
Public Policy / 公开政策
Non-code book content is publicly licensed under CC BY-NC-SA 4.0 unless a file says otherwise.
非代码书籍内容默认按 CC BY-NC-SA 4.0 公开授权，除非文件另有说明。
Third-party commercial use requires separate permission from LifeBook Shufang and relevant rights holders.
第三方商业使用必须另行取得 LifeBook 书坊及相关权利人的授权。
Contributor rules are defined in license/CONTRIBUTING*.md.
Select the matching language-pair template, for example French-to-English, Japanese-to-Spanish, English-to-Simplified-Chinese, or Traditional-Chinese-to-German.
选择匹配的语言方向模板，例如 French-to-English、Japanese-to-Spanish、English-to-Simplified-Chinese 或 Traditional-Chinese-to-German。
If a matching target-language framework exists under template/epub_pipeline/targets/{target}/, read it before translating.
如果 template/epub_pipeline/targets/{target}/ 下存在匹配的目标语言质量框架，翻译前必须读取。
Create the book project with books/scripts/create_book_project.py; it copies template/epub_pipeline/common into books/{target}/{number}_{book_id_slug}/, then overlays the matching language-pair template.
使用 books/scripts/create_book_project.py 创建书籍工程；脚本会把 template/epub_pipeline/common 复制到 books/{target}/{number}_{book_id_slug}/，再覆盖复制匹配的语言方向模板。
Write all book-specific source text, translations, QA, EPUB output, and metadata only inside the book project.
所有具体书籍的原文、译文、QA、EPUB 输出和 metadata 只能写入书籍工程目录。
Run the workflow through source evidence, rights checks, research, trial translation, chapter translation, review, gates, EPUB production, validation, mandatory post-EPUB stratified random spot-check, independent review, versioned release, and retrospective.
按来源证据、版权核查、研究、试译、章节翻译、审校、门禁、EPUB 制作、校验、EPUB 后强制分层随机抽检、独立评审、版本化发布和复盘流程执行。
For translation, chapter review, random spot-check, reader feedback, and final artifact review, use skills/expert-translation-quality/SKILL.md to enforce expert-level prose and downstream-context polysemy back-checking without bloating every prompt.
翻译、章节审校、随机抽检、读者反馈和最终产物审阅时，使用 skills/expert-translation-quality/SKILL.md，把专家级译文与后文回看多义词复查作为执行流程，而不是把长规则塞进每个 prompt。
Before EPUB production, run node scripts/publication_lint.js --target={target-language} --write-report inside the book project and fix all hard errors.
EPUB 制作前，必须在书籍工程内运行 node scripts/publication_lint.js --target={target-language} --write-report，并修复所有硬错误。
Before a translated or final chapter can pass chapter controls or template preflight, also run the structural translation coverage gate: npm run check:translation-coverage or python scripts/check_translation_coverage.py --write-report. This gate compares chapters/src/ against chapters/translated/ and chapters/final/ for severe paragraph-block shrinkage, chapter character shrinkage, lost footnote references, lost footnote definitions, and lost table/figure/formula units. A chapter control file that claims PASS cannot override this machine gate.
已译章节或终稿章节通过章节控制或模板预检前，还必须运行结构性译文覆盖门禁：npm run check:translation-coverage 或 python scripts/check_translation_coverage.py --write-report。该门禁对比 chapters/src/ 与 chapters/translated/、chapters/final/，拦截严重段落块缩水、章节字数异常缩水、脚注引用丢失、脚注定义丢失，以及表格/图片/公式单元丢失。章节 control 自称 PASS 不能覆盖这个机器门禁。
Before final EPUB output, check long chapter titles against references/chapter_title_policy.md and any matching source-to-target title strategy. EPUB navigation must use concise labels, not printed-TOC title chains.
最终 EPUB 输出前，必须按 references/chapter_title_policy.md 及对应语言方向标题策略检查长章节标题。EPUB 目录应使用短题名，不应塞入纸书目录式标题链。
After the first full-book EPUB is built, run the stratified random spot-check module from references/stratified_random_spotcheck.md and prompts/16a_stratified_random_spotcheck.md. Use deterministic scripts, keep all samples and evidence under books/{target}/{number}_{book_id_slug}/reviews/random_spotcheck/round_XXX/, close every discovered P0/P1/P2, and run npm run review:random-validate:pass before final output.
第一版全书 EPUB 生成后，必须执行 references/stratified_random_spotcheck.md 和 prompts/16a_stratified_random_spotcheck.md 定义的分层随机抽检模块。必须使用确定性脚本，所有样本和证据保存在 books/{target}/{number}_{book_id_slug}/reviews/random_spotcheck/round_XXX/，所有发现的 P0/P1/P2 必须关闭，最终输出前必须通过 npm run review:random-validate:pass。
The current executor must generate new random spot-check rounds. Historical PASS rounds from previous agents, previous releases, or previous private artifacts do not count. The user may specify any current-run consecutive PASS requirement of >=1; when unspecified, the default validator requires the latest consecutive 2 PASS rounds from the same review_run_id.
当前执行者必须生成新的随机抽检轮次。之前 Agent、之前 release 或之前 private artifact 已经 PASS 的历史轮次不得计入本次执行。用户可以指定任意 >=1 的当前运行连续 PASS 轮次要求；未指定时，默认强校验要求同一 review_run_id 下最新连续 2 轮 PASS。
After random spot-check closure, run the versioned release module from references/release_versioning.md and prompts/18a_release_versioning.md. Create output/release/book_vX.X.X.epub, release_note_vX.X.X.md, release_state.json, and release_index.md; output/book.epub alone is not enough for DONE.
随机抽检闭环通过后，必须执行 references/release_versioning.md 和 prompts/18a_release_versioning.md 定义的版本化发布模块。必须创建 output/release/book_vX.X.X.epub、release_note_vX.X.X.md、release_state.json 和 release_index.md；只有 output/book.epub 不能标记 DONE。
For translated titles and important proper nouns, title occurrences do not count as first body occurrences. Chapter titles, subtitles, and navigation titles must follow target-language title style; source names or parenthetical original names belong at the first natural body occurrence, in notes, or in glossary/proper_nouns.csv. If the user does not specify a proper-noun display setting, use policy 3: first natural body occurrence target（source）, later target. Note markers are a separate hard rule: only [1], (1), fullwidth （1）, or 注1 marker families are allowed.
对译文标题和重点专有名词而言，标题中的出现不计入“正文首次出现”。章节标题、副标题和目录题名必须按目标语言标题习惯处理；原文名或括注原名应放在正文第一次自然出现处、注释或 glossary/proper_nouns.csv 中。用户未指定专名显示设置时，使用策略 3：第一次正文自然出现 译名（原文），后续用译名。注号是独立硬规则：只能使用 [1]、(1)、全角 （1） 或 注1 体系。
If a specific book has systematic refinement issues, place its goal under books/{target}/{number}_{book_id_slug}/goal/, then backfill reusable lessons into common, target-language, or language-pair templates.
如果某本书有系统性精修问题，目标文档应放在 books/{target}/{number}_{book_id_slug}/goal/，再把可复用经验回填到 common、目标语言或语言方向模板。
Put language-pair-specific scripts, datasets, and exploratory research under research/{language-pair-template}/... or the matching language-pair template, not in the repository root.
特定语言方向的脚本、数据集和探索性调研应放在 research/{language-pair-template}/... 或对应语言方向模板中，不要放在仓库根目录。
Do not hard-code local absolute paths in scripts or prompts. Resolve paths from the script location, the repository root, or explicit user-provided arguments.
不要在脚本或 prompt 中写死本机绝对路径。路径应基于脚本位置、仓库根目录或用户显式传入的参数解析。
