md-document

Imported from borghei/claude-skills (markdown-html/md-document/SKILL.md). Install upstream with npx skills add borghei/claude-skills --skill md-document. Copyright stays with the author (MIT + Commons Clause).

Markdown Document Publishing

Turn an authored markdown file into a single HTML document you can email, host, or print — semantic structure, an automatic table of contents, figures and tables numbered and referenceable by number, footnotes, and a print stylesheet that survives contact with a PDF exporter. One file out, no runtime dependencies, no external assets.

When to use this skill

• Publishing a report or whitepaper that must arrive as one file, not a folder
• Producing a PDF from markdown without a LaTeX or Pandoc toolchain
• Numbering figures and tables so prose can reference them instead of saying "below"
• Converting untrusted or contributed markdown where injection safety matters
• Standardizing a document series so every issue looks like the same publication
• Catching broken cross-references before a document reaches readers

Inputs the skill expects

• A markdown source file, optionally with YAML frontmatter
• Figure images as relative paths, or as base64 data URIs (PNG, JPEG, GIF, WebP or AVIF) for true single-file output
• The intended output: screen, print, or both
• Document length and whether a table of contents is warranted
• Page geometry, if printing: size, margins, single- or double-sided
• A stylesheet, if not using the bundled theme

Clarify First

Before converting, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.

• Screen, print, or both: why it changes the output: print needs a page profile, forced light colors, and break control; skipping it produces a document that looks right on screen and breaks on paper.
• Whether images must be embedded: why it changes the output: relative image paths mean the HTML file is not actually self-contained, which defeats the point if it will be emailed.
• Document length and TOC expectation: why it changes the output: a TOC on a two-page memo is noise; depth 3 on a long report produces a TOC longer than some sections.
• Whether the markdown is trusted: why it changes the output: it does not change escaping (always on), but it determines whether a review gate should run first.

Ask about the two or three inputs that matter most for this request, not the whole list. If the user wants a result straight away, produce it and state your assumptions at the top.

Workflows

Workflow 1 — Convert a document to self-contained HTML

1. Audit labels and references first — the converter's gate reports broken references, but the auditor explains what to do about each one.
2. Convert. The bundled theme is inlined automatically; pass --css to override.
3. Check the gate. A non-zero exit means the rendered document contains visible [?fig:name] markers where numbers should be.

python3 markdown-html/md-document/scripts/crossref_auditor.py \
  --input markdown-html/md-document/assets/sample_document.md --format text

python3 markdown-html/md-document/scripts/md_to_html.py \
  --input markdown-html/md-document/assets/sample_document.md \
  --out build/report.html --toc-depth 2 --format text

Workflow 2 — Produce a print-ready PDF

1. Generate the print block from a page profile and append it to the theme, so the exported file carries its own print rules.
2. Convert with the extended stylesheet.
3. Open in a browser, print to PDF with margins set to Default — a browser margin setting overrides @page and will clip content.

python3 markdown-html/md-document/scripts/print_profile.py \
  --input markdown-html/md-document/assets/sample_print_profile.json \
  --out build/print.css --format text

cat markdown-html/md-document/assets/document_theme.css build/print.css > build/full.css

python3 markdown-html/md-document/scripts/md_to_html.py \
  --input markdown-html/md-document/assets/sample_document.md \
  --css build/full.css --out build/report.html

Workflow 3 — Gate a document series in CI

1. Run the auditor at warning severity so unlabelled figures block, not just broken references.
2. Run the conversion gate; it fails on unresolved references and undefined footnotes — both render as visible defects.
3. Emit JSON for both so a CI job can annotate the diff.

python3 markdown-html/md-document/scripts/crossref_auditor.py \
  --input markdown-html/md-document/assets/sample_document.md --max-severity warning --format json

python3 markdown-html/md-document/scripts/md_to_html.py \
  --input markdown-html/md-document/assets/sample_document.md --out build/report.html --format json

Decision frameworks

Table-of-contents depth

If the TOC exceeds one screen, reduce the depth. A contents list longer than the first section is a navigation failure, not thoroughness.

Reference style

[PROVEN] Never use positional language in a document that may be paginated. "Below" breaks when the table lands on the next page, breaks silently when a section is reordered, and means nothing to a reader navigating by heading.

Alt text versus caption

The auditor flags placeholder alt text (chart, image, screenshot, empty) at error severity and alt text under 15 or over 125 characters at warning.

Print geometry

At A4 with 20mm margins the text column is ~92 characters — well outside the 55-85 comfort band. Widening the margins is the fix; max-width: none on main is what causes the problem.

Anti-Patterns

The "see the table below" reference

Mistake: Writing positional prose — "the chart below", "as shown above" — instead of a numbered cross-reference. Why it happens: It reads naturally while drafting, when the author can see the whole document at once and the table genuinely is below. Instead: Write [@tbl:policies]. Pagination moves content, reordering breaks positional claims silently, and a reader navigating by heading has no "below". The auditor cannot detect a broken "below"; it fails the build on a broken [@tbl:policies].

Allowing raw HTML through the converter

Mistake: Adding an escape hatch so authors can drop <div class="..."> or an embed into the markdown. Why it happens: A real formatting need appears that the subset does not cover, and passing HTML through is a one-line change. Instead: Extend the subset or the stylesheet. The escape-then-render ordering is the entire security model — the moment raw HTML passes through, every document becomes an injection vector, and the converter can no longer be pointed at contributed content. There is deliberately no --allow-html flag.

Reusing the caption as alt text

Mistake: Writing one string and letting it serve as both the figure caption and the alt attribute. Why it happens: The converter falls back to exactly this when no caption is given, which makes it look sanctioned. Instead: Write both. The caption tells a sighted reader what to conclude; the alt text describes what the figure shows to someone who cannot see it. "Figure 3. Costs fall 40% under Policy B" is a fine caption and useless alt text — it states the conclusion without describing the chart.

Print rules added after the fact

Mistake: Building the document for screen, then bolting on a print stylesheet when someone asks for a PDF. Why it happens: Print feels like a rendering detail rather than a design constraint, and the screen version already looks finished. Instead: Decide print-or-not before converting. Retrofitted print CSS produces the classic failures — stranded headings, tables split mid-row, dark theme reaching paper as invisible gray text, a 92-character measure. print_profile.py exists so the geometry is a reviewed input, not an afterthought.

Trusting the gate as a proof of quality

Mistake: Treating a passing conversion as evidence the document is ready to publish. Why it happens: The gate is automated and green, which reads as authoritative. Instead: The gate checks that references resolve and footnotes are defined. It cannot see a stranded heading, a figure separated from its caption, a table split across pages, or a PDF whose margins clipped the content. Proof every page of the actual output at 100% zoom before publishing.

Files

All scripts share one exit-code contract: 0 clean, 2 gate failed (findings at or above the threshold), 1 the tool itself errored. A CI job can therefore tell a real defect from a broken invocation.

Three CLI tools, one module. md_render.py is a library, not a fourth command — it holds the parser that md_to_html.py imports. A single-file converter came to 429 lines, well over the 300-line ceiling. Splitting CLI from parser is the remedy the tool-design standard prescribes for an oversized script, and same-directory imports keep the package self-contained: md-slides carries its own separate slide_render.py rather than importing this one.

Scope & Limitations

In Scope:

• Converts one markdown file into a self-contained HTML document with the theme inlined, a table of contents, numbered figures and tables, [@fig:], [@tbl:] and [@sec:] cross-references, and footnotes (md_to_html.py).
• Audits labels, references, alt text and heading hierarchy before conversion (crossref_auditor.py).
• Generates a print/PDF stylesheet from a JSON page profile covering page size, margins, mirrored margins and running heads (print_profile.py).
• Escapes all source text, including raw HTML, and passes link and image URLs through a scheme allowlist.

Out of Scope:

• Producing the PDF file itself. The skill writes print CSS; the PDF comes from a browser's print dialog.
• Slide decks; see markdown-html/md-slides/.
• Pre-publication review of links, readability, terminology and frontmatter; see markdown-html/md-review/.
• Building or validating theme tokens and contrast; see markdown-html/design-system/.
• Reading or auditing existing Word or PDF files; see documents/docx-toolkit/ and documents/pdf-toolkit/.

Important Caveats:

• The converter supports a subset of markdown. Nested lists, reference-style links and raw HTML are not rendered; anything outside the subset appears as escaped text. A URL with a literal parenthesis truncates unless it is percent-encoded.
• The converter does not embed image files itself. A document with relative image paths needs those files beside it. For a single file, write the image as a base64 data URI (data:image/png;base64,...); the allowlist accepts PNG, JPEG, GIF, WebP and AVIF data URIs in an image source only. SVG data URIs, other data: types and data: link targets are replaced with #blocked-url.
• Fenced code gets a language-* class but no syntax highlighting.
• The gate checks only that references resolve and footnotes are defined. Stranded headings, split tables and clipped margins show up only when the printed output is proofed, and browsers differ in how much of @page they honour.