---
name: aperivue/render-pdf-doc
source: https://app.decimal.ai/s/aperivue-render-pdf-doc@1/SKILL.md
source_sha256: f2c93e4bb4c8
---

# Render-PDF-Doc Skill

Markdown + frontmatter → publication-quality academic PDF (English or Korean).

## Why This Skill Exists

In real circulation cycles for academic PDFs, two recurring failure patterns appear:
1. v1 drafts: change-history, version numbers, and PI attribution leak into the attached PDF, confusing the first recipient.
2. v2 drafts: pandoc pipe-table dash ratios are misjudged, narrowing the first column and forcing label wrapping that hurts readability.

Manual fixes work but the same pattern recurs across proposals, briefings, IRB covers, exemption applications. This skill focuses on **layout** (CJK fonts + table column widths). Bibliography and CSL are handled by `/manage-refs`.

## Boundary (separation from other skills)

| Task | Skill |
|---|---|
| Manuscript + bibliography → DOCX/PDF | `/manage-refs scripts/render_pandoc.sh` (CSL + .bib) |
| Filling an institutional .docx form | `/fill-protocol` |
| ICMJE COI form | `/fill-icmje-coi` |
| Figure / PPTX | `/make-figures`, `/present-paper` |
| **This skill**: non-bib academic markdown → PDF (proposal, briefing, anchor doc, IRB cover) | `/render-pdf-doc` |

## Core Principles

1. **Pipe table column widths must be inferred from content.** No equal splitting. Size the first column (label) to the longest label, and distribute the remaining width content-proportionally across the data columns.
2. **Set the CJK font explicitly** — `mainfont` + `CJKmainfont`. The default fallback is OS-detected.
3. **For circulation PDFs, remove change history / version numbers / PI attribution** (or split them into a supplementary). Use the frontmatter `redact_internal: true` option.
4. **No Quarto dependency** — raw pandoc + xelatex. Quarto's `tbl-colwidths` has reported PDF regressions (issues 6089/9200).

## Dependencies

```bash
# macOS
brew install pandoc
brew install --cask mactex-no-gui          # xelatex + xeCJK (~5 GB)

# Linux
sudo apt-get install pandoc texlive-xetex texlive-lang-cjk fonts-noto-cjk

# Windows (PowerShell) — run in Git Bash afterwards
winget install --id JohnMacFarlane.Pandoc
winget install --id MiKTeX.MiKTeX          # xelatex; installs missing LaTeX packages on demand
# No CJK font download needed: Malgun Gothic ships with Windows 7+ and is the default here.
```

Detection:
```bash
bash scripts/check_deps.sh
```

**Windows / Git Bash note.** MiKTeX's binary directory
(`%LOCALAPPDATA%\Programs\MiKTeX\miktex\bin\x64`) is often not on the Git Bash `PATH`,
so `xelatex` can read as `[MISS]` even after install. Both `check_deps.sh` and
`render_pdf.sh` now auto-probe that location; if `xelatex` still isn't found, add the
directory to your `PATH` (or run from the *MiKTeX Console → Settings*-configured shell).
The Windows CJK/main font default is **Malgun Gothic** (preinstalled); override per document
via frontmatter, or with `--font` / `--cjk-font`.

## Workflow

### Step 1 — Author markdown with frontmatter

```yaml
---
title: "Paper 2 Calibration Anchor — Q&A Grid"
author: "<Author Group>"
date: "2026-05-01"
mainfont: "Apple SD Gothic Neo"        # macOS default
CJKmainfont: "Apple SD Gothic Neo"
geometry: "margin=0.85in"
fontsize: 11pt
linestretch: 1.25
colorlinks: true
---
```

For Linux/CI, use `Noto Sans CJK KR`; on Windows, use `Malgun Gothic`. The render script auto-detects the default per OS.

### Step 2 — Infer column widths

```bash
python scripts/infer_colwidths.py input.md > input.colwidths.md
```

The script:
1. Finds every pipe table block.
2. For each column, computes display width = `max(len(header), max(len(cell)))` (CJK = 2 cells, ASCII = 1).
3. Generates dash-row separator with proportional dash counts.
4. Writes a new file with separator rows replaced.

Override per-table via attribute: `{tbl-colwidths="[20,40,40]"}` after caption — passes through unchanged.

### Step 3 — Render

```bash
bash scripts/render_pdf.sh -i input.colwidths.md -o output.pdf
```

Or one-shot:
```bash
bash scripts/render_pdf.sh -i input.md -o output.pdf --infer-colwidths
```

### Step 3.5 — Scientific-symbol + CJK glyph scan (before render)

xelatex **silently drops** any character the chosen font does not cover — the PDF
renders with the glyph simply missing, no error or warning. Academic markdown
routinely carries glyphs a default Latin font misses: transition arrows (→ ↑ ↓),
math operators (− ≤ ≥ ± √ ∪ × ≈ ≠), stats Greek (κ μ σ β), bullets/marks (• ★ ✓),
and CJK. Scan the source first so a silent drop is caught before it ships:

```bash
python3 scripts/scan_glyph_coverage.py input.md --strict
# real cmap check when you have the font file + fonttools:
python3 scripts/scan_glyph_coverage.py input.md --font "/path/to/body.otf" --strict
```

It groups the risky glyphs by class (advisory), or — with `--font` + `fonttools`
— reports which are genuinely absent from the font's cmap. If risky glyphs are
present, ensure `mainfont`/`CJKmainfont` cover them (a CJK-capable font such as
*Apple SD Gothic Neo* / *Noto Sans CJK* usually covers arrows + Hangul but can
still miss the true-minus `−` U+2212 and `★`). **The DOCX is authoritative; the
PDF is a convenience copy** — never let a PDF render drop a glyph the document
needs.

### Step 4 — Visual verify

Open the PDF. Check:
- The first-column labels do not wrap and stay on a single line
- Data columns have sufficient width
- No broken Korean glyphs (a Times New Roman fallback means CJKmainfont was not applied)
- No missing scientific symbols (arrows, −, ≤, ±, √) — the Step 3.5 scan flags candidates
- No change history / internal version numbers exposed

## Templates

Starter markdown in `templates/` (English default; a Korean variant `*_ko.md` ships alongside each):
- `anchor-doc.md` — Q&A grid
- `proposal-cover.md` — research-proposal cover page
- `briefing-handout.md` — meeting brief (1-page)
- `reference-table.md` — comparison-table format

Each template marks slots with a `<!-- TODO: -->` marker.

## Anti-Patterns

| Anti-pattern | Consequence |
|---|---|
| Equal dash split (`\|---\|---\|---\|`) | A column with only a short label gets the same width → cramped data columns |
| `CJKmainfont` not set | Hangul falls back to Times New Roman (broken Latin glyphs or blanks) |
| Change history / version (e.g. v3.2.2) / PI attribution exposed in a circulation PDF | Confuses the first recipient; leaks internal information |
| Quarto `tbl-colwidths` for PDF | PDF regression in Quarto 1.4+ — trust HTML only |

## Files

- `scripts/render_pdf.sh` — pandoc + xelatex wrapper, OS font detection
- `scripts/infer_colwidths.py` — auto-generates pipe-table separator dash ratios
- `scripts/check_deps.sh` — checks for pandoc / xelatex / CJK font
- `templates/` — 4 starters (English) + their `*_ko.md` Korean variants
- `references/pandoc_korean_cheatsheet.md` — collection of frontmatter patterns (Korean-PDF reference)
- `references/known_pitfalls.md` — em-dash line breaks, smart quotes, etc. (Korean-PDF reference)

## Anti-Hallucination

- Numerical content in tables: read every value from the source CSV or analysis output. Do not
  retype a number from prose, and do not carry one forward from an earlier draft — a table that
  was correct in v3 is not evidence it is correct in v4.
- References: use `/manage-refs` separately — this skill does not handle bib.
- When producing a circulation PDF, keep the received primary source (the `.docx` or `.eml` a
  co-author sent) unmodified as its own artifact. The PDF is a derivative, not a replacement, and
  the next round is diffed against that source. If the source was itself AI-drafted by a
  collaborator, treat every number, denominator, and author-year in it as unverified: re-derive
  each from the underlying paper or analysis output before it reaches the PDF.

## Global-rule references

Some passages in this skill cite a path of the form `~/.claude/rules/<name>.md`. Those are the
maintainer's personal global rules, kept outside this repository. They are **not shipped with
this skill** and will not exist on your machine; they appear only as provenance for where a
convention came from. If one of them looks like it is standing in for an instruction you actually
need, that is a bug — please open an issue, because the instruction belongs here.