lov-any2pdf skill

PDF大师 · PDF Master

This skill converts any Markdown file into a publication-quality PDF using Python's reportlab library. It was developed through extensive iteration on real Chinese technical reports and solves several hard problems that naive MD→PDF converters get wrong.

Triggers

Activate when

• User wants to convert .md → .pdf
• User has a markdown report/document and wants professional typesetting
• Document contains CJK characters (Chinese/Japanese/Korean) mixed with Latin text
• Document has fenced code blocks, markdown tables, or nested lists
• Document has local/remote images, Obsidian callouts, emoji, or math formulas
• User wants a cover page, table of contents, or watermark in their PDF

Do not activate when

• The user only wants to revise Markdown content and does not need a PDF output.
• The source is HTML, DOCX, or another format and the user has not authorized a Markdown conversion step.
• The user needs a slide deck, editable Word document, or image export rather than a PDF.

Quick Start

python md2pdf/scripts/md2pdf.py \
  --input report.md \
  --output report.pdf \
  --title "My Report" \
  --author "Author Name"

All parameters except --input are optional. When no theme is supplied, the renderer selects one from the document's content and structure.

Configuration Resolution (MANDATORY)

Do not present a full configuration form before every conversion. Resolve each setting independently in this order:

1. the user's explicit instruction in the current request, then CLI arguments or document frontmatter;
2. the user's last explicitly confirmed value from the lov_any2pdf preference namespace;
3. a content-aware decision for the theme, or the documented safe default for non-visual settings.

Proceed without asking when this resolves the conversion. Ask at most one concise question only when an unresolved choice would materially change the result, such as a missing local asset path or permission to generate a new image. Do not ask the user to choose from every theme unless they explicitly request the theme catalog.

After the user explicitly confirms or changes a reusable parameter, record that value as the latest preference. Reuse it next time when the user is silent. Never persist an inferred value as though the user confirmed it, and honor requests to reset or ignore saved preferences.

Reusable parameters include theme, cover/TOC, page size, frontispiece mode, watermark text and geometry, image-cover mode, back-cover mode/banner, disclaimer, copyright, header, and footer. Revalidate saved file paths before use; if an asset no longer exists, ask for a replacement instead of silently substituting one.

Content-Aware Theme Decision

When neither the current request nor the saved preferences specify a theme, read the title, opening, headings, document length, code blocks, tables, imagery, and publishing context, then select the theme that best serves the article. The following are decision cues, not a fixed keyword router:

Content character	Good starting theme
Long Chinese prose, knowledge notes, dense tables	songti-reading
Strategy, business, market, or research report	consulting-navy
Academic paper with abstract/method/results/references	ieee-journal or classic-thesis
Code-heavy technical guide or API documentation	github-light
Data narrative with charts and metrics	tufte
Literary, art, photography, or design essay	ink-wash
Formal Chinese policy or institutional document	chinese-red
Neutral English prose or print-first document	paper-classic

warm-academic remains available when its tone genuinely fits the article, but it is never the default merely because the user omitted a style.

Mapping User Choices to CLI Args

Choice	CLI arg
Design style	--theme with value from table below (latex-clean uses pandoc engine)
Frontispiece local	--frontispiece <path>
Frontispiece AI	Generate image first, then --frontispiece /tmp/frontispiece.png
Watermark text	--watermark "文字"
Watermark style	--wm-size 30 --wm-opacity 0.1 --wm-angle 45 (all optional)
Image as cover	--image-cover true (requires --frontispiece)
Back cover image	--banner <path>
Back cover text	--disclaimer "声明" and/or --copyright "© 信息"

Theme Name Mapping

Style	--theme value	Inspiration
暖学术	warm-academic	Warm editorial-academic palette
经典论文	classic-thesis	LaTeX classicthesis
Tufte	tufte	Edward Tufte's books
期刊蓝	ieee-journal	IEEE journal format
精装书	elegant-book	LaTeX ElegantBook
中国红	chinese-red	Chinese formal documents
水墨	ink-wash	水墨画 / ink wash painting
GitHub	github-light	GitHub Markdown style
Nord	nord-frost	Nord color scheme
海洋	ocean-breeze	Clean teal editorial palette
LaTeX 清爽	latex-clean	pandoc+XeLaTeX 原生排版，无封面
咨询深蓝	consulting-navy	Consulting research reports
宋黑阅读	songti-reading	中文出版物常见的宋体正文 + 黑体标题层级

Handling AI-Generated Frontispiece

If user chose AI generation: read the document title + first paragraphs, use an image generation tool to create a themed illustration matching the chosen design style, show for approval, then pass via --frontispiece /path/to/image.png

Architecture

Markdown → Preprocess (split merged headings) → Parse (code-fence-aware) → Story (reportlab flowables) → PDF build

Key components:

1. Font system: Palatino (Latin body), Songti SC Regular (CJK body), selectable serif/sans CJK emphasis, Menlo (code) on macOS; auto-fallback on Linux/Windows
2. CJK wrapper: _font_wrap() wraps CJK character runs in <font> tags for automatic font switching
3. Mixed text renderer: _draw_mixed() handles CJK/Latin mixed text on canvas (cover, headers, footers)
4. Code block handler: esc_code() preserves indentation, mid-line alignment, and line breaks in reportlab Paragraphs
5. Smart table widths: Proportional column widths based on content length, with 18mm minimum
6. Bookmark system: ChapterMark flowable creates PDF sidebar bookmarks + named anchors
7. Heading preprocessor: _preprocess_md() splits merged headings like # Part## Chapter into separate lines
8. Image handler: local, relative, file://, and remote markdown images are scaled into the body frame with fallback text on errors
9. Callout renderer: Obsidian-style > [!NOTE] blocks render as themed boxed callouts
10. Formula renderer: display formulas use optional matplotlib mathtext images, with styled text fallback
11. Emoji fallback: emoji are rendered as cached Twemoji PNGs when available, or with a local emoji font fallback

Hard-Won Lessons

CJK Characters Rendering as □

reportlab's Paragraph only uses the font in ParagraphStyle. If fontName="Mono" but text contains Chinese, they render as □. Fix: Always apply _font_wrap() to ALL text that might contain CJK, including code blocks.

Code Blocks Losing Line Breaks

reportlab treats \n as whitespace. Fix: esc_code() converts \n → <br/> and all spaces →  , preserving indentation and mid-line alignment before _font_wrap().

CJK/Latin Word Wrapping

Default reportlab breaks lines only at spaces, causing ugly splits like "Claude\nCode". Fix: Set wordWrap='CJK' on body/bullet styles to allow breaks at CJK character boundaries.

Songti Reading Needs a Sans Hierarchy

Using one sans family everywhere gives dense Chinese reports a uniform dark texture, while using Songti everywhere weakens headings and table headers. Fix: use the songti-reading theme for Songti SC Regular body text, Palatino Latin prose, sans-serif Chinese headings/emphasis, Menlo code, wider leading, and serif table bodies. Keep this as an explicit theme rather than silently changing every report.

On macOS, Songti.ttc index 0 is Songti SC Black, not Regular. The readable face is index 6. Common report symbols such as ✓, ✗, and → need an explicit symbol-font fallback or they may render as □ even when the surrounding Chinese is correct.

Canvas Text with CJK (Cover/Footer)

drawString() / drawCentredString() with a Latin font can't render 年/月/日 etc. Fix: Use _draw_mixed() for ALL user-content canvas text (dates, stats, disclaimers).

Images Silently Dropped (Relative Paths)

An image reference to charts/chart_01.png in a markdown file used to get skipped without warning because the image path was resolved against the current working directory, not the markdown's directory. Fix: main() now passes input_dir (the .md's directory) into the builder, and the image handler resolves relative paths against it. Missing images now also emit a WARN: image not found: ... to stderr instead of silently dropping.

Multi-Line Image References (pandoc --wrap=auto)

When feeding pandoc's output into md2pdf, pandoc's default --wrap=auto (72 cols) wraps long image references to paths such as path.png across multiple lines, which defeated the single-line image regex. Fix: _preprocess_md() now collapses multi-line image references into one line (outside code fences) before parsing.

Pipeline tip: If you're piping HTML → markdown via pandoc, use pandoc --wrap=none input.html -o output.md to avoid wrap-related parsing issues for images and tables alike.

Input Format

This skill takes Markdown files only as input. If you have HTML, DOCX, or other formats, convert them to markdown first (e.g. pandoc --wrap=none). Embedded HTML blocks in markdown are passed through as text — pre-process any visual content (charts, complex tables) into plain markdown tables or image references before invoking md2pdf.

Configuration Reference

Most options can also be set in top-of-file YAML-style frontmatter. Explicit CLI arguments take precedence over frontmatter values.

Argument	Frontmatter Key	Default	Description
--input	—	(required)	Path to markdown file
--output	—	output.pdf	Output PDF path
--title	title	From first H1	Document title for cover page
--subtitle	subtitle	""	Subtitle text
--author	author	""	Author name
--date	date	Today	Date string
--version	version	""	Version string for cover
--watermark	watermark	""	Watermark text (empty = none)
--theme	theme	content-aware auto	Color theme name; explicit and saved preferences take precedence
--theme-file	—	""	Custom theme JSON file path
--cover	cover	true	Generate cover page
--toc	toc	true	Generate table of contents
--page-size	page-size	A4	Page size (A4 or Letter)
--frontispiece	frontispiece	""	Full-page image after cover
--banner	banner	bundled assets/backcover-banner.jpg	Back cover banner image (pass none to disable)
--header-title	header-title	""	Report title in page header
--footer-left	footer-left	author	Brand/author in footer
--stats-line	stats-line	""	Stats on cover
--stats-line2	stats-line2	""	Second stats line
--edition-line	edition-line	""	Edition line at cover bottom
--disclaimer	disclaimer	""	Back cover disclaimer
--copyright	copyright	""	Back cover copyright
--code-max-lines	code-max-lines	30	Max lines per code block
--image-cover	image-cover	false	Use frontispiece as full-bleed cover (page 1), text cover becomes page 2
--heading-top-spacer	heading-top-spacer	5	Top spacer before H1/H2 chapter titles in mm
--wm-size	wm-size	auto (half page width)	Watermark font size; auto-scales so text width ≈ 50% of page width
--wm-opacity	wm-opacity	0.1	Watermark opacity (0.0–1.0)
--wm-angle	wm-angle	35	Watermark rotation angle in degrees
--wm-spacing-x	wm-spacing-x	9999	Watermark horizontal spacing in pt (≥2000 = single centered per page)
--wm-spacing-y	wm-spacing-y	9999	Watermark vertical spacing in pt (≥2000 = single centered per page)

Themes

Available: songti-reading, warm-academic, nord-frost, github-light, solarized-light, paper-classic, ocean-breeze, monokai-warm, dracula-soft, tufte, classic-thesis, ieee-journal, elegant-book, chinese-red, ink-wash, and consulting-navy.

Each theme defines: page background, ink color, accent color, faded text, border, code background, watermark tint.

Fallback: pandoc + XeLaTeX

When reportlab fails (e.g. wide/complex tables cause infinite loops in table width calculation), fall back to pandoc with XeLaTeX. This produces high-quality CJK PDFs with proper table handling.

When to use pandoc engine

• User chose latex-clean theme (k) — pandoc is the primary engine, not a fallback
• Document has many wide multi-column tables (reportlab's table layout may hang)
• Document needs LaTeX-quality typesetting (justified text, hyphenation)
• reportlab md2pdf.py hangs or crashes on the input

Basic command

pandoc input.md -o output.pdf \
  --pdf-engine=xelatex \
  -V CJKmainfont="Songti SC" -V mainfont="Palatino" -V monofont="Menlo" \
  -V geometry:margin=2.5cm -V fontsize=11pt \
  --toc -V toc-title="目录" -V documentclass=article

Adding watermark + headers/footers

pandoc input.md -o output.pdf \
  --pdf-engine=xelatex \
  -V CJKmainfont="Songti SC" -V mainfont="Palatino" -V monofont="Menlo" \
  -V geometry:margin=2.5cm -V fontsize=11pt \
  -V colorlinks=true -V linkcolor=red -V toccolor=red -V urlcolor=red \
  --toc -V toc-title="目录" -V documentclass=article \
  -V header-includes='
\usepackage{fancyhdr}
\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\small 页眉左侧文字}
\fancyhead[R]{\small 页眉右侧文字}
\fancyfoot[C]{\thepage}
\usepackage{draftwatermark}
\SetWatermarkText{水印文字}
\SetWatermarkScale{0.5}
\SetWatermarkColor[gray]{0.9}
'

Pandoc theme presets

Each preset defines a complete set of pandoc -V flags. Use the full command from "Adding watermark + headers/footers" above, replacing the color/link flags per preset.

Theme	linkcolor	toccolor	urlcolor	Watermark color	Notes
chinese-red	red	red	red	[gray]{0.9}	朱红正式，适合政企报告、白皮书
warm-academic	brown	brown	brown	[gray]{0.9}	陶土色调，温润学术风
classic-thesis	brown	brown	brown	[gray]{0.85}	LaTeX classicthesis 灵感
ieee-journal	blue	blue	blue	[gray]{0.9}	藏蓝严谨，期刊风格
github-light	blue	blue	blue	[gray]{0.92}	极简蓝白，程序员友好
ink-wash	black	black	black	[gray]{0.92}	水墨素雅，文学/设计类
nord-frost	teal	teal	teal	[gray]{0.9}	北欧冰霜蓝灰
consulting-navy	blue	blue	blue	[gray]{0.88}	咨询深蓝，研究报告风

chinese-red 完整示例（本次一滕项目实际使用）

pandoc input.md -o output.pdf \
  --pdf-engine=xelatex \
  -V CJKmainfont="Songti SC" -V mainfont="Palatino" -V monofont="Menlo" \
  -V geometry:margin=2.5cm -V fontsize=11pt \
  -V colorlinks=true -V linkcolor=red -V toccolor=red -V urlcolor=red \
  --toc -V toc-title="目录" -V documentclass=article \
  -V header-includes='
\usepackage{fancyhdr}
\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\small 报告标题 | 品牌名}
\fancyhead[R]{\small 商业机密}
\fancyfoot[C]{\thepage}
\usepackage{draftwatermark}
\SetWatermarkText{商业机密}
\SetWatermarkScale{0.5}
\SetWatermarkColor[gray]{0.9}
'

latex-clean 完整示例（最简洁的 LaTeX 学术风格）

选择 latex-clean 时，跳过 reportlab，直接使用 pandoc 生成。特点：无封面、无扉页、 无装饰色块，纯 LaTeX 排版 + 点线 TOC 引导符，适合内容为王的学术/技术文档。

pandoc input.md -o output.pdf \
  --pdf-engine=xelatex \
  -V CJKmainfont="Songti SC" -V mainfont="Palatino" -V monofont="Menlo" \
  -V geometry:margin=2.5cm -V fontsize=11pt \
  -V colorlinks=true -V linkcolor=brown -V toccolor=brown -V urlcolor=brown \
  --toc -V toc-title="目录" -V documentclass=article

如需添加页眉页脚或水印，参考上方 "Adding watermark + headers/footers" 追加 -V header-includes。

Known limitations (pandoc fallback)

• No cover page (pandoc article class has no built-in cover — use --include-before-body with a LaTeX snippet if needed)
• Frontispiece/back cover not supported (use md2pdf.py for these)
• → ★ ☆ symbols may warn in Palatino — they render via CJK font fallback, safe to ignore
• ASCII art diagrams render as code blocks (same as md2pdf.py)

Dependencies (pandoc fallback)

Requires pandoc and a TeX distribution with XeLaTeX:

brew install pandoc
brew install --cask mactex-no-gui   # or basictex

Dependencies

pip install reportlab --break-system-packages
# Optional formula rendering:
pip install matplotlib --break-system-packages

Recommended Ubuntu/Debian fonts:

sudo apt install fonts-dejavu-core fonts-liberation fonts-freefont-ttf fonts-noto fonts-noto-cjk fonts-noto-color-emoji

维护备注

• 默认 layout 对每个 H2 强制 PageBreak（md2pdf.py:1899），中文大纲类文档会产出大量半空页；在自定义主题的 layout 里加 "continuous_headings": true（songti-reading 用的就是这个）（2026-09-05, 1c67671）
• 正文里的 ASCII 流程图会按普通文本断词换行；靠调小 code_size/code_leading（如 10.0→8.6、16.2→14.2）让整行放得下，不要改图（2026-09-05, 1c67671）
• 实际脚本路径是 scripts/md2pdf.py，SKILL.md 正文写的 md2pdf/scripts/md2pdf.py 不存在（2026-09-05, 1c67671）

Runtime context (shared)

运行前读取本 Skill 包的 skill.yaml，由宿主提供 skill-runtime/v1 上下文。字段解析顺序为：当前请求、项目上下文、个人 Preferences、品牌 Profile、通用默认值。

• 只使用 Manifest 声明的字段；Profile 保存公开品牌事实，Preferences 保存个人工作偏好。
• required: true 字段缺失时，按 Manifest 的问题配置向用户提出一个聚焦问题；用户明确同意后再保存回答。
• 报错提供可复制的 context_id、字段路径与来源，诊断内容避开秘密、完整私人路径和原始配置。

通用反馈闭环

用户在 Skill 驱动任务中提出修改意见时，继续当前产物前必须执行：

1. 先判断意见是 task-specific（仅本次）还是 reusable（可跨任务复用）。
2. task-specific 只修改当前任务，不改 Skill。
3. reusable 先确定作用域：领域规则先更新对应 canonical Skill；适用于所有 Skill 的规则先更新共享规范。
4. 完成规则更新、版本、lint 与分发核验后，再把修改应用到当前任务。
5. reusable 修改会使此前的“确认”“继续”“发吧”失效；完成当前产物修改和回读后必须停下，等待用户下一步指示，不自动进入发布、提交或其他外部写入。