Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Write academic papers in Markdown with Pandoc for multi-format output
.claude/skills/brycewang-stanford-markdown-academic-guide/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 139% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 35% | 0% |
| case-03 | ✓→✓ | = Same ✓ | 64% | 0% |
| case-04 | ✓→✓ | = Same ✓ | 84% | 0% |
| case-05 | ✓→✓ | = Same ✓ | 63% | 0% |
A skill for writing academic papers in plain-text Markdown and converting them to PDF, Word, LaTeX, and HTML using Pandoc. Covers YAML metadata, citation management, cross-references, templates, and workflows for collaborative academic writing.
1. Plain text: Version-controllable with Git (diff-friendly)
2. Portable: Works on any OS, any editor
3. Pandoc: Convert to PDF, DOCX, LaTeX, HTML, EPUB
4. Focus: Content-first writing without formatting distractions
5. Citations: Pandoc-citeproc handles bibliography automatically
6. Collaboration: Easy to review diffs in pull requestsUse Markdown when:
- You need multi-format output (PDF + Word + HTML)
- Collaborators prefer Word but you prefer plain text
- The paper has standard formatting needs
- You want a simpler syntax than LaTeX
Use LaTeX directly when:
- The journal provides a mandatory LaTeX template
- You need advanced typesetting (complex math layouts, custom floats)
- You are writing a thesis with institutional LaTeX requirementsyaml--- title: "Your Paper Title: A Markdown-Based Approach" author: - name: Jane Smith affiliation: Department of Computer Science, University X email: jane@university.edu orcid: 0000-0002-1234-5678 - name: John Doe affiliation: School of Engineering, University Y date: 2026-03-09 abstract: | This paper demonstrates how academic manuscripts can be written in plain Markdown and converted to publication-quality documents using Pandoc. We show that this approach reduces formatting overhead while maintaining full citation and cross-reference capabilities. keywords: [academic writing, Markdown, Pandoc, reproducible research] bibliography: references.bib csl: apa-7th-edition.csl link-citations: true numbersections: true ---
markdown# Introduction Academic writing often involves tedious formatting tasks that distract from content creation [@smith2024; @jones2023, pp. 45-50]. Recent tools enable plain-text workflows that separate content from presentation [see @garcia2022, chap. 3]. ## Background As @lee2021 demonstrated, Markdown-based workflows reduce formatting errors by 40% compared to WYSIWYG editors. ### Subsection Example Inline math: $E = mc^2$ Display math: $$ \hat{\beta} = (X^T X)^{-1} X^T y $$
[@key] -> (Author, 2024)
@key -> Author (2024)
[@key, p. 42] -> (Author, 2024, p. 42)
[@key1; @key2] -> (Author1, 2024; Author2, 2023)
[-@key] -> (2024) -- suppress author name
[see @key] -> (see Author, 2024)bash# Markdown to PDF (via LaTeX) pandoc paper.md -o paper.pdf \ --citeproc \ --number-sections \ --pdf-engine=xelatex # Markdown to Word (DOCX) pandoc paper.md -o paper.docx \ --citeproc \ --reference-doc=template.docx # Markdown to LaTeX pandoc paper.md -o paper.tex \ --citeproc \ --standalone # Markdown to HTML pandoc paper.md -o paper.html \ --citeproc \ --standalone \ --mathjax
bash# Extract default template for customization pandoc -D latex > custom-template.tex # Use custom template pandoc paper.md -o paper.pdf \ --template=custom-template.tex \ --citeproc \ --pdf-engine=xelatex
markdownSee @fig:architecture for the system overview. {#fig:architecture width=80%} Results are shown in @tbl:results. | Method | Accuracy | F1 Score | |--------|----------|----------| | Ours | 0.95 | 0.93 | | Baseline| 0.88 | 0.85 | : Comparison of methods. {#tbl:results} As proven in @eq:main, the relationship holds. $$y = \alpha + \beta x + \epsilon$$ {#eq:main}
bash# Compile with cross-references pandoc paper.md -o paper.pdf \ --filter pandoc-crossref \ --citeproc \ --pdf-engine=xelatex
pythondef markdown_collaboration_workflow() -> dict: """ Recommended workflow for multi-author Markdown papers. """ return { "setup": [ "Create a Git repository for the paper", "Add .gitignore for PDF output and LaTeX aux files", "Store references.bib in the repo", "Include a Makefile for reproducible builds" ], "writing": [ "Each author works on a branch", "Use pull requests for section drafts", "Review diffs in GitHub/GitLab (plain text diffs are readable)", "Resolve merge conflicts in plain text (much easier than .docx)" ], "makefile_example": ( "all: paper.pdf paper.docx\n" "paper.pdf: paper.md references.bib\n" "\tpandoc paper.md -o paper.pdf --citeproc --pdf-engine=xelatex\n" "paper.docx: paper.md references.bib\n" "\tpandoc paper.md -o paper.docx --citeproc --reference-doc=template.docx\n" "clean:\n" "\trm -f paper.pdf paper.docx" ) }
Pandoc handles most academic writing needs, but has limitations with complex table layouts, advanced figure placement, and journal-specific LaTeX class features. For final submission, you may need to fine-tune the generated LaTeX or DOCX output. Keep your Markdown source as the canonical version and treat generated files as disposable build artifacts.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 12,498 | 40,412 | +223% | 1 | 1 | 0% | 1,891 | 3,104 | +64% | 0 | 0 | — |
case-02 | fail→pass | 76,656 | 5,462 | -93% | 1 | 1 | 0% | 1,082 | 2,588 | +139% | 0 | 0 | — |
case-03 | pass→pass | 7,281 | 5,314 | -27% | 1 | 1 | 0% | 1,439 | 2,363 | +64% | 0 | 0 | — |
case-04 | pass→pass | 39,694 | 5,479 | -86% | 1 | 1 | 0% | 1,367 | 2,519 | +84% | 0 | 0 | — |
case-05 | pass→pass | 7,703 | 3,771 | -51% | 1 | 1 | 0% | 1,367 | 2,235 | +63% | 0 | 0 | — |
case-06 | pass→pass | 5,709 | 3,633 | -36% | 1 | 1 | 0% | 1,075 | 2,290 | +113% | 0 | 0 | — |
case-07 | pass→pass | 10,011 | 3,463 | -65% | 1 | 1 | 0% | 1,532 | 2,209 | +44% | 0 | 0 | — |
case-08 | fail→pass | 12,525 | 6,467 | -48% | 1 | 1 | 0% | 1,884 | 2,541 | +35% | 0 | 0 | — |
case-09 | pass→pass | 8,289 | 5,579 | -33% | 1 | 1 | 0% | 1,529 | 2,375 | +55% | 0 | 0 | — |
case-10 | pass→pass | 10,186 | 10,247 | +1% | 1 | 1 | 0% | 1,469 | 2,883 | +96% | 0 | 0 | — |
case-11 | pass→pass | 6,638 | 3,354 | -49% | 1 | 1 | 0% | 1,194 | 2,232 | +87% | 0 | 0 | — |
case-12 | pass→pass | 5,969 | 3,206 | -46% | 1 | 1 | 0% | 1,009 | 2,135 | +112% | 0 | 0 | — |
case-13 | pass→pass | 7,567 | 4,908 | -35% | 1 | 1 | 0% | 1,132 | 2,300 | +103% | 0 | 0 | — |
case-14 | pass→pass | 5,374 | 3,513 | -35% | 1 | 1 | 0% | 748 | 2,235 | +199% | 0 | 0 | — |
case-15 | pass→pass | 7,655 | 2,695 | -65% | 1 | 1 | 0% | 1,202 | 2,092 | +74% | 0 | 0 | — |
case-16 | pass→pass | 5,627 | 3,593 | -36% | 1 | 1 | 0% | 988 | 2,253 | +128% | 0 | 0 | — |
case-17 | pass→pass | 16,315 | 13,670 | -16% | 1 | 1 | 0% | 2,668 | 3,957 | +48% | 0 | 0 | — |
case-18 | pass→pass | 20,833 | 19,996 | -4% | 1 | 1 | 0% | 2,910 | 4,591 | +58% | 0 | 0 | — |
case-19 | pass→pass | 14,657 | 11,202 | -24% | 1 | 1 | 0% | 2,190 | 3,256 | +49% | 0 | 0 | — |
case-20 | pass→pass | 14,004 | 12,095 | -14% | 1 | 1 | 0% | 2,272 | 3,607 | +59% | 0 | 0 | — |
case-21 | pass→pass | 11,632 | 18,764 | +61% | 1 | 1 | 0% | 2,150 | 4,415 | +105% | 0 | 0 | — |
case-22 | pass→pass | 20,123 | 19,789 | -2% | 1 | 1 | 0% | 3,878 | 5,725 | +48% | 0 | 0 | — |
case-23 | pass→pass | 19,003 | 18,047 | -5% | 1 | 1 | 0% | 3,526 | 4,892 | +39% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 22 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +9 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.