Latex Manuscript

LaTeX Manuscript Handling

Patterns for reducing friction when working in LaTeX manuscript repos.

1. Repo reconnaissance

LaTeX manuscripts follow predictable layouts. Run this upfront:

# Find the master .tex file (the one with \documentclass)
grep -rl '\\documentclass' *.tex

# Enumerate all \input / \include to get the section tree
grep -rn '\\input\|\\include' MASTER.tex

# Find all figure directories
find . -name 'figure*' -type d

# Find custom macros / variables
grep -rl '\\newcommand\|\\DeclareFigureNumber\|\\def' *.tex other/*.tex

Read the variables/macros file early — it defines figure reference commands, custom labels, and numbering that you need to parse the rest of the manuscript.

2. Viewing PDF figures

PDF figures can't be viewed directly. Convert to PNG first:

# macOS (built-in, fast, no dependencies)
sips -s format png figure.pdf --out figure.png

# Linux (requires poppler-utils)
pdftoppm -png -r 200 -singlefile figure.pdf figure

Batch convert all figures in a repo:

for d in figures/figure_*/; do
  [ -f "$d/figure.pdf" ] && sips -s format png "$d/figure.pdf" --out "$d/figure.png"
done

Then use the Read tool on the .png files to visually inspect them.

3. Reading the manuscript efficiently

Parallel section reads: Launch multiple Read calls simultaneously for all section files. A typical structure:

• text/abstract.tex, text/introduction.tex, text/results.tex, text/discussion.tex, text/methods.tex
• other/title.tex, other/variables.tex

Figure legends: Each figure directory typically has legend.tex. Read all legends in one batch with a subagent.

Methods can be huge (500-1000+ lines). Read in chunks or use Grep to find specific sections by \subsection or \label.

4. Finding issues in manuscript text

# Placeholder text and editorial notes
grep -rn 'TODO\|FIXME\|XXX\|figure X\|fig X\|(stats)\|BS suggested\|PLACEHOLDER' text/

# Overly precise p-values (might need rounding)
grep -rn 'p=[0-9]\.[0-9]*e-' text/

# Check for "novel" / overclaiming language
grep -rn 'novel\|first\|unprecedented\|breakthrough' text/

5. LaTeX editing style

When editing LaTeX manuscripts:

• Match the existing style exactly: indentation, spacing, citation format (\autocite vs \cite), figure reference macros
• Check for custom commands before writing raw LaTeX. E.g., the repo may define \figref{key} instead of raw \textbf{Figure N}
• Preserve line structure: Many LaTeX repos put one sentence or one paragraph per line. Don't reflow unless asked
• Use the Edit tool, not sed/awk. The Edit tool shows exactly what changes and preserves surrounding context
• Test that edits don't break compilation by checking for unmatched braces, unclosed environments, etc.

Common patterns:

% Adding a statistical test result
(paired $t$-test, $p < 0.05$)
(Wilcoxon signed-rank test, $p = 0.0078$)
($p < 0.001$, permutation test, 1000 shuffles)

% Adding a limitation acknowledgment
We note that this comparison was performed on $N=4$ animals; replication in larger cohorts would further strengthen these conclusions.

% Softening a claim
% BEFORE: "FR provides a novel and flexible approach"
% AFTER:  "FR provides an effective approach"

6. Working with reviewer response documents (.docx)

Reviewer response docs are usually Word files. Use the docx skill's unpack/edit/pack workflow:

# Unpack to editable XML
python3 scripts/office/unpack.py responses.docx unpacked/

# Edit unpacked/word/document.xml with the Edit tool

# Repack
python3 scripts/office/pack.py unpacked/ responses_revised.docx --original responses.docx

Extracting text from reviewer XML for analysis:

import xml.etree.ElementTree as ET
tree = ET.parse('unpacked/word/document.xml')
ns = {'w': 'http://schemas.openxmlformats.org/wordprocessingml/2006/main'}
for p in tree.getroot().iter(f'{{{ns["w"]}}}p'):
    texts = []
    for t in p.iter(f'{{{ns["w"]}}}t'):
        if t.text: texts.append(t.text)
    line = ''.join(texts).strip()
    if line:
        bold = p.find('.//w:rPr/w:b[@w:val="1"]', ns)
        gray = p.find('.//w:rPr/w:color[@w:val="999999"]', ns)
        tag = '[REVIEWER]' if gray else '[RESPONSE]' if bold else ''
        print(f'{tag} {line[:200]}')

Formatting conventions in response docs:

• Reviewer text is typically gray (color val="999999")
• Author responses are bold
• TODOs / incomplete items are often red (color val="ff0000")
• When inserting new response paragraphs into the XML, copy the <w:pPr> and <w:rPr> structure from existing response paragraphs to match formatting

7. Subagent strategy

For manuscript work, subagents are most useful for:

• Parallel reads: Read all section files + all legends simultaneously
• Figure conversion: Batch convert PDFs to PNGs
• Targeted edits: Send a subagent to make a specific set of edits to one file while you work on another
• Legend/methods audits: "Check all figure legends for missing error bar descriptions"

Avoid using subagents for decisions that require cross-section understanding (e.g., "should we move this figure to the supplement?"). Do those yourself.

8. Common reviewer response patterns

Checklist of things reviewers commonly request that can be addressed via manuscript text alone: