Book From Transcripts

Book From Transcripts

Output Goal

Produce a reader-facing book package, not a production note dump:

• 完整正稿.md: title, preface/summary, table of contents, chapters, glossary, sources.
• chapters/*.md: clean chapter files.
• 出版版式终稿.docx and optional 出版版式终稿.pdf.
• manifest.*: source URLs, dates, transcript availability, and processing status.
• *.zip: only final reader-facing artifacts unless the user asks for raw working files.

Never include raw copyrighted full transcripts in final shareable outputs unless the user explicitly owns or supplies them and asks for inclusion. Keep fetched transcripts in work/ by default.

Workflow

1. Set scope. Identify topic, source set, audience, target language, output format, and whether the user wants an internal reading draft or a publication-like final. If the user says “直接开始”, make reasonable assumptions and proceed.
2. Build a source manifest. Record title, URL, source type, date if available, transcript path or status, rights note, and planned chapter/batch. For public sources, cite links; for private/local source files, cite filenames.
3. Create the book architecture. Group sources into parts/chapters by narrative logic, not by raw episode boundaries when that hurts readability. Draft a book title, subtitle, chapter titles, and a one-page 序 that tells readers what the book covers.
4. Translate and edit. Preserve technical facts, names, company terms, and numbers. Use domain Chinese that reads like a business/technology book, not subtitles. For detailed rules, read references/editorial-standards.md.
5. Assemble Markdown. Create a master manuscript with this order:
   • # 书名：副标题
   • ## 序
   • ## 目录
   • chapters as # 第N章 ...
   • # 术语与人物表
   • # 来源说明
6. Clean the draft. Run scripts/scrub_book_markdown.py on the master manuscript and inspect its report. Remove duplicated generated headings, prompt residue, chunk markers, “给你看的说明”, “版权信息”, and other process text unless the user explicitly wants them.
7. Build DOCX/PDF. Prefer the documents skill for Word creation and render QA. If a deterministic local builder is useful, run scripts/build_book_docx.py on the clean Markdown, then render the DOCX with the Documents skill renderer.
8. Verify visually and structurally. Render to PNG/PDF, inspect front matter, chapter starts, chapter endings, glossary, and source pages. Run blank-page and residue checks. Iterate until clean.
9. Deliver only final artifacts. Link final files from outputs/. Mention material limitations briefly only when relevant, e.g. source authorization or missing transcript.

Translation And Editing

When translating with DeepSeek, OpenAI, or another model:

• Chunk long sources, but assemble by chapter and remove chunk seams.
• Ask the model for faithful, field-appropriate Chinese, not simplification.
• Preserve English product/company names when standard in Chinese tech writing: Google, YouTube, Chrome, Android, OpenAI, Transformer, TPU.
• Add a glossary for recurring technical terms, founders, products, and acquisitions.
• Use chapter summaries to create book-level coherence.
• Do not leave “以下是…”, “本稿用于…”, “根据转写稿…”, prompt fragments, model apologies, or transcript labels in the final manuscript.

Layout Guidance

Default book layout:

• A5 page size, generous margins, Chinese Songti body, restrained Heiti headings.
• Clean cover with title, subtitle, year or edition only.
• A reader-facing 序 before the table of contents.
• No copyright/legal page unless the user asks.
• Running header/footer in body pages; no header/footer on cover.
• Sources and glossary in appendices.

For detailed book QA and typography rules, read references/production-checklist.md.

Useful Scripts

• scripts/scrub_book_markdown.py input.md --out clean.md --report report.txt Cleans common generation residue and reports suspicious lines.
• scripts/build_book_docx.py clean.md --out book.docx --title "..." --subtitle "..." Builds a deterministic A5 DOCX from clean Markdown. Use when no custom document builder already exists.

Run scripts with the bundled workspace Python when available. If the task already uses the documents skill, use its renderer for final DOCX visual QA.

Done Criteria

Before final response:

• Final package includes Markdown, DOCX, PDF if requested, chapters, and manifest/source list.
• Front matter reads like a real book: cover -> sequence/preface -> TOC -> Chapter 1.
• No raw transcript files in shareable output unless explicitly requested.
• No duplicated chapter headings, truncation, chunk seams, model/prompt residue, or empty pages.
• DOCX metadata is scrubbed when possible.
• PDF/PNG render has been inspected at least on cover, preface, TOC, all chapter starts, low-ink pages, glossary, and source page.