paper-compile — Claude Code skill: what it does, how to install
Compiles LaTeX papers to PDF using latexmk, diagnoses build errors, and verifies page limits for academic conferences.
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a claude-codeWhat paper-compile does
The skill runs latexmk with pdflatex, xelatex, or lualatex and automatically parses compile.log. If errors occur, it enters an iterative loop with up to 3 fix attempts. It resolves missing packages via tlmgr, fixes mismatched BibTeX syntax, repairs broken figure paths between .pdf and .png, and removes leftover [VERIFY] markers. If compilation is still stuck after 2 attempts and the Codex plugin is present, it can trigger /codex:rescue for an independent check.
Once the PDF compiles, the skill checks file integrity: the file must exceed 100KB, with no ?? or [?] markers from missing labels. An automated check with pdftotext inspects the page budget. It distinguishes between ML conference standards (where references do not count against the limit for NeurIPS or ICLR) and IEEE venues where references are strictly counted. If the paper overflows, it highlights bloated sections and recommends moving specific proofs to the appendix.
This skill requires local installations of latexmk, pdflatex, and bibtex, meaning builds fail immediately if TeX Live is absent.
What it helps with
- Resolve repetitive LaTeX build errors and missing package failures
- Validate conference page limits for ML venues or IEEE journals
- Detect broken references, missing citations, and unresolved labels
- Spot figure path mismatches and leftover verification markers
Example prompts for the agent
- compile paper in paper/ directory and check the page count
- build PDF for paper/main.tex and fix any missing citations
- run paper-compile on paper/ and check for undefined references
How to install paper-compile
1Skills CLI (Vercel)
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a claude-code2Manually: put the files in place
~/.claude/skills/paper-compile/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/paper-compile/
1Skills CLI (Vercel)
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a codex2Manually: put the files in place
.agents/skills/paper-compile/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a cursor2Manually: put the files in place
.cursor/skills/paper-compile/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install wanshuiyin/Auto-claude-code-research-in-sleep paper-compile2Manually: put the files in place
.github/skills/paper-compile/SKILL.md1Skills CLI (Vercel)
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a gemini-cli2Platform CLI
gemini skills install https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep --path skills/paper-compileManually: put the files in place
.devin/skills/paper-compile/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about paper-compile
How to install paper-compile in Claude Code?
Run npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill paper-compile -a claude-code. This command registers the skill directly into Claude Code.
What does paper-compile do?
It builds LaTeX papers via latexmk, inspects compile.log for errors, applies automated code fixes up to 3 times, and verifies page count rules.
What host dependencies are required?
You need latexmk, pdflatex, and bibtex installed in your PATH. Section and page count verification also relies on pdftotext and Python 3.
Source text
SKILL.md · MIT
---
name: paper-compile
description: "Compile LaTeX paper to PDF, fix errors, and verify output. Use when user says \"编译论文\", \"compile paper\", \"build PDF\", \"生成PDF\", or wants to compile LaTeX into a submission-ready PDF."
argument-hint: "[paper-directory]"
allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob
---
# Paper Compile: LaTeX to Submission-Ready PDF
Compile the LaTeX paper and fix any issues: **$ARGUMENTS**
## Constants
- **COMPILER = `latexmk`** — LaTeX build tool. Handles multi-pass compilation automatically.
- **ENGINE = `pdflatex`** — LaTeX engine. Options: `pdflatex` (default), `xelatex` (for CJK/custom fonts), `lualatex`.
- **MAX_COMPILE_ATTEMPTS = 3** — Maximum attempts to fix errors and recompile.
- **PAPER_DIR = `paper/`** — Directory containing LaTeX source files.
- **MAX_PAGES** — Page limit. ML conferences: main body to Conclusion end (excluding references & appendix). ICLR=9, NeurIPS=9, ICML=8. **IEEE venues: references ARE included in page count.** IEEE journal ≈ 12-14 pages, IEEE conference ≈ 5-8 pages (all inclusive).
## Workflow
### Step 1: Verify Prerequisites
Check that the compilation environment is ready:
```bash
# Check LaTeX installation
which pdflatex && which latexmk && which bibtex
# If not installed, provide instructions:
# macOS: brew install --cask mactex-no-gui
# Ubuntu: sudo apt-get install texlive-full
# Server: conda install -c conda-forge texlive-core
```
Verify all required files exist:
```bash
# Must exist
ls $PAPER_DIR/main.tex
# Should exist
ls $PAPER_DIR/references.bib
ls $PAPER_DIR/sections/*.tex
ls $PAPER_DIR/figures/*.pdf 2>/dev/null || ls $PAPER_DIR/figures/*.png 2>/dev/null
```
### Step 2: First Compilation Attempt
```bash
cd $PAPER_DIR
# Clean previous build artifacts
latexmk -C
# Full compilation (pdflatex + bibtex + pdflatex × 2)
latexmk -pdf -interaction=nonstopmode -halt-on-error main.tex 2>&1 | tee compile.log
```
### Step 3: Error Diagnosis and Auto-Fix
If compilation fails, read `compile.log` and fix common errors:
**Missing packages:**
```
! LaTeX Error: File `somepackage.sty' not found.
```
→ Install via `tlmgr install somepackage` or remove the `\usepackage` if unused.
**Undefined references:**
```
LaTeX Warning: Reference `fig:xyz' on page 3 undefined
```
→ Check `\label{fig:xyz}` exists in the correct figure environment.
**Missing figures:**
```
! LaTeX Error: File `figures/fig1.pdf' not found.
```
→ Check if the file exists with a different extension (.png vs .pdf). Update the `\includegraphics` path.
**Citation undefined:**
```
LaTeX Warning: Citation `smith2024' undefined
```
→ Add the missing entry to `references.bib` or fix the citation key.
**`[VERIFY]` markers in text:**
→ Search for `[VERIFY]` markers left by `/paper-write`. These indicate unverified citations or facts. Search for the correct information or flag to the user.
**Overfull hbox:**
```
Overfull \hbox (12.5pt too wide) in paragraph at lines 42--45
```
→ Minor: usually ignorable. If severe (>20pt), rephrase the text or adjust figure width.
**BibTeX errors:**
```
I was expecting a `,' or a `}'---line 15 of references.bib
```
→ Fix BibTeX syntax (missing comma, unmatched braces, special characters in title).
**`\crefname` undefined for custom theorem types:**
→ Ensure `\crefname{assumption}{Assumption}{Assumptions}` and similar are in the preamble after `\newtheorem{assumption}`.
### Step 4: Iterative Fix Loop
```
for attempt in 1..MAX_COMPILE_ATTEMPTS:
compile()
if success:
break
parse_errors()
auto_fix()
```
For each error:
1. Read the error message from `compile.log`
2. Locate the source file and line number
3. Apply the fix
4. Recompile
**Stuck after 2 attempts?** If Codex plugin is installed, invoke `/codex:rescue` — Codex can independently read the LaTeX source and `compile.log` to spot issues Claude missed (e.g., conflicting packages, encoding problems, subtle macro errors). If not installed, continue with Claude's own diagnosis.
### Step 5: Post-Compilation Checks
After successful compilation, verify the output:
```bash
# Check PDF exists and has content
ls -la main.pdf
# Check page count
pdfinfo main.pdf | grep Pages
# macOS: open for visual inspection
# open main.pdf
```
**Visual review (automated):**
If the compiled PDF exists, read it directly to check visual presentation:
- Figure quality: readable labels, legible text, distinguishable colors
- Layout: no orphaned section headers, no awkward page breaks
- Figures appear near their first text reference (not pages away)
- Tables: aligned columns, consistent decimal precision
- No overfull content visibly extending past margins
This is a quick visual scan, not a full review — the improvement loop does deeper visual review.
**Automated checks:**
- [ ] PDF file exists and is > 100KB (not empty/corrupt)
- [ ] Total page count is reasonable (MAX_PAGES + appendix + references)
- [ ] No "??" in the PDF (undefined references — grep the log)
- [ ] No "[?]" in the PDF (undefined citations — grep the log)
- [ ] Figures are rendered (not missing image placeholders)
```bash
# Check for undefined references
grep -c "LaTeX Warning.*undefined" compile.log
# Check for missing citations
grep -c "Citation.*undefined" compile.log
```
### Step 6: Page Count Verification
**CRITICAL**: Verify paper fits within MAX_PAGES.
**For ML conferences (ICLR/NeurIPS/ICML/CVPR/ACL/AAAI):** Main body = first page through end of Conclusion section (not necessarily §5 — could be §6, §7, or §8 depending on structure). References and appendix are NOT counted.
**For IEEE venues:** The TOTAL page count (including references) must fit within the limit. There is no separate "main body" counting — everything up to and including the references counts.
**Precise check using `pdftotext`:**
```bash
# Extract text and find where Conclusion ends vs References begin
pdftotext main.pdf - | python3 -c "
import sys
text = sys.stdin.read()
pages = text.split('\f')
for i, page in enumerate(pages):
if 'Ethics Statement' in page or 'Reproducibility' in page:
print(f'Conclusion ends on page {i+1}')
if any(w in page for w in ['References', 'Bibliography']):
lines = [l for l in page.split('\n') if l.strip()]
for l in lines[:3]:
if 'References' in l or 'Bibliography' in l:
print(f'References start on page {i+1}')
break
"
```
If Conclusion ends mid-page and References start on the same page, the main body is that page number (e.g., if both are on page 9, main body = ~8.5 pages, which is fine for a 9-page limit since it leaves room for the References header).
If over limit:
- Identify which sections are longest
- Suggest specific cuts (move proofs to appendix, compress tables, tighten writing)
- Report: "Main body is X pages (limit: MAX_PAGES). Suggestion: move [specific content] to appendix."
### Step 6.5: Stale File Detection
Check for orphaned section files not referenced by `main.tex`:
```bash
# Find all .tex files in sections/ and check which are \input'ed by main.tex
for f in paper/sections/*.tex; do
base=$(basename "$f")
if ! grep -q "$base" paper/main.tex; then
echo "WARNING: $f is not referenced by main.tex — consider removing"
fi
done
```
This prevents confusion from leftover files when section structure changes (e.g., old `5_conclusion.tex` left behind after restructuring to 7 sections).
### Step 7: Submission Readiness
For conference submission, additional checks:
- [ ] **Anonymous**: no author names, affiliations, or self-citations that reveal identity
- [ ] **Page limit**: main body within MAX_PAGES (to end of Conclusion)
- [ ] **Font embedding**: all fonts embedded in PDF
```bash
pdffonts main.pdf | grep -v "yes" # should return nothing (or only header)
```
- [ ] **No supplementary mixed in**: appendix clearly after `\newpage\appendix`
- [ ] **File size**: reasonable (< 50MB for most venues, < 10MB preferred)
- [ ] **No `[VERIFY]` markers**: search the PDF text for leftover markers
### Step 8: Output Summary
```markdown
## Compilation Report
- **Status**: SUCCESS / FAILED
- **PDF**: paper/main.pdf
- **Pages**: X (main body to Conclusion) + Y (references) + Z (appendix)
- **Within page limit**: YES/NO (MAX_PAGES = N)
- **Errors fixed**: [list of auto-fixed issues]
- **Warnings remaining**: [list of non-critical warnings]
- **Undefined references**: 0
- **Undefined citations**: 0
### Next Steps
- [ ] Visual inspection of PDF
- [ ] Run `/paper-write` to fix any content issues
- [ ] Submit to [venue] via OpenReview / CMT / HotCRP
```
## Key Rules
- **Never delete the user's source files** — only modify to fix errors
- **Keep compile.log** — useful for debugging
- **Don't suppress warnings** — report them, let the user decide
- **If LaTeX is not installed**, provide clear installation instructions rather than failing silently
- **Font embedding is critical** — some venues reject PDFs with non-embedded fonts
- **Page count rules differ by venue** — ML conferences: main body to Conclusion (refs excluded). **IEEE venues: total pages including references.**
## Common Venue Requirements
| Venue | Style File | Citation | Page Limit | Refs in limit? | Submission |
|-------|-----------|----------|------------|----------------|------------|
| ICLR 2026 | `iclr2026_conference.sty` | `natbib` (`\citep`/`\citet`) | 9 pages (to Conclusion end) | No | OpenReview |
| NeurIPS 2025 | `neurips_2025.sty` | `natbib` (`\citep`/`\citet`) | 9 pages (to Conclusion end) | No | OpenReview |
| ICML 2025 | `icml2025.sty` | `natbib` (`\citep`/`\citet`) | 8 pages (to Conclusion end) | No | OpenReview |
| IEEE Journal | `IEEEtran.cls` [journal] | `cite` (`\cite{}`, numeric) | ~12-14 pages (Transactions) / ~4-5 (Letters) | **Yes** | IEEE Author Portal / ScholarOne |
| IEEE Conference | `IEEEtran.cls` [conference] | `cite` (`\cite{}`, numeric) | 5-8 pages (varies by conf) | **Yes** | EDAS / IEEE Author Portal |
© wanshuiyin · MIT · wanshuiyin/Auto-claude-code-research-in-sleep