paper-latex-writing

Trackable LaTeX academic paper editing with session-based logging

Purpose

Perform systematic, trackable edits to LaTeX academic papers using coding-by-logging methodology. All changes are logged with before/after diffs, metrics tracking, and git-compatible version control.

Capabilities

1. Length Reduction

•Identify verbose sections and redundant content
•Cut to target page/word count while preserving key points
•Iterative reduction with metrics tracking
•Section-level or global scope

2. Clarity Improvement

•Rewrite dense paragraphs for readability
•Simplify complex sentences
•Improve logical flow and transitions
•Strengthen topic sentences

3. Citation Management

•Fix broken/missing citations
•Add citations for uncited claims
•Remove unnecessary citations
•Ensure consistent citation style
•Update bibliography entries

4. Structural Reorganization

•Move sections/subsections
•Merge related content
•Split overly long sections
•Reorder for better narrative flow

5. Figure/Table Management

•Fix broken references
•Add/update captions
•Ensure all figures are referenced
•Check figure placement and sizing

6. Consistency Checking

•Terminology consistency (e.g., "dataset" vs "data set")
•Notation consistency (math symbols, variable names)
•Formatting consistency (capitalization, hyphenation)
•Style guide compliance (e.g., ICML, NeurIPS)

Paper Structure Auto-Detection

Hybrid Detection Strategy

On first invocation, the skill automatically:

•

Scans current directory for paper structure:

code

- Main .tex file: Find file with \documentclass
- Bibliography: Find .bib file
- Sections: Detect \input{} patterns
- Structure type: Modular (separate files) or Monolithic (single file)
- Logging location: paper/code-logging/ or create if missing

•

Creates .paper-latex-config.yaml with detected structure:

yaml

# Auto-generated on 2026-01-17
main_tex: fairglucose_icml2026.tex
bibliography: fairglucose_icml2026.bib
sections_dir: _sections/
structure: modular
logging_dir: paper/code-logging/

# Section mapping (auto-detected from \input commands)
sections:
  abstract: _sections/00_abstract.tex
  introduction: _sections/01_introduction.tex
  related_work: _sections/02_related_work.tex
  dataset_design: _sections/03_dataset_design.tex
  benchmark_protocol: _sections/04_benchmark_protocol.tex
  baseline_results: _sections/05_baseline_results.tex
  discussion: _sections/06_discussion.tex
  conclusion: _sections/07_conclusion.tex

appendix_sections:
  dataset_statistics: _sections/appendix/A_dataset_statistics.tex
  model_evaluation: _sections/appendix/B_model_evaluation.tex
  experiment_details: _sections/appendix/C_experiment_details.tex

# Figures and tables
figures_dir: Figures/
tables_dir: Table/
appendix_figures_dir: AppendixFigure/
appendix_tables_dir: AppendixTable/

# Compilation
compiler: pdflatex
bibtex: bibtex

# Target metrics (can be updated)
target_pages: 8
current_pages: 35
venue: ICML 2026
track: Datasets & Benchmarks

•
User can override by editing .paper-latex-config.yaml or providing custom config:
bash
```
/paper-latex-writing --config custom-paper-config.yaml "reduce to 10 pages"
```

Fallback Behavior

If auto-detection fails:

•Ask user to specify main .tex file
•Prompt for structure type (modular/monolithic)
•Create minimal config and proceed

Session-Based Logging

Log File Structure

code

paper/code-logging/
├── latex-writing-sessions.md      # Main log (all sessions)
└── archives/
    ├── session-001-length-reduction.md
    ├── session-002-clarity-pass.md
    └── session-003-citation-fixes.md

Log Format

CRITICAL: Inline Modification Format

All changes in session logs MUST use the strikethrough/CC-modify notation for trackability:

Format Rules:

code

**P1**: ~~Old content description~~
        New content description >> CC-modify

**P2**: ~~Text to delete~~ >> CC-modify: DELETE

**P3 (NEW)**: New paragraph to add >> CC-modify

Notation Standards:

•KEEP = no comment (leave original text as-is)
•MODIFY = ~~old~~ on first line, new text on second line (indented with spaces) >> CC-modify
•DELETE (single line) = ~~text~~ >> CC-modify: DELETE
•DELETE (multi-line) = Each line gets ~~ markers, then >> CC-modify at end
•ADD = new text >> CC-modify (mark as NEW)

Readability tip: For long modifications, always use two lines:

•Line 1: ~~strikethrough old text~~
•Line 2: Indented new text (4+ spaces) >> CC-modify

Multi-line deletion example:

code

~~**Subsection Title**:~~
~~First paragraph to remove...~~
~~Second paragraph to remove...~~
>> CC-modify: DELETE

LaTeX-specific examples (with % comments for compilation safety):

When editing .tex files directly, ALL tracking notation must be commented with %:

latex

% ~~\subsection{Verbose Title That's Too Long}~~
%     \subsection{Concise Title} >> CC-modify
\subsection{Concise Title}

% ~~Figure~\ref{fig:removed} shows the detailed analysis...~~
% >> CC-modify: DELETE

% \textbf{Event Impact} (NEW): Events increase error by 8.2 mg/dL. >> CC-modify
\textbf{Event Impact}: Events increase error by 8.2 mg/dL.

CRITICAL LaTeX Rule:

•All ~~strikethrough~~ lines → prefix with %
•All >> CC-modify markers → prefix with %
•Actual new code → NO % (must compile)
•This keeps tracking visible in source but doesn't break compilation

Complete LaTeX modification example:

latex

% File: _sections/05_baseline_results.tex
% Lines: 45-52

% ~~\subsection{Subgroup Performance Analysis}~~
%     \subsection{Subgroup Performance and Generalization} >> CC-modify
\subsection{Subgroup Performance and Generalization}

% ~~Across all demographic and clinical subgroups, we observe substantial~~
% ~~heterogeneity in model performance. Male patients in the 18-39 age group~~
% ~~consistently demonstrate lower prediction error compared to their female~~
% ~~counterparts...~~
%     Model performance varies substantially across demographic strata:
%     \begin{itemize}
%     \item \textbf{Age-Gender}: Males 18-39 outperform females by 3.2 mg/dL
%     \item \textbf{Diabetes Type}: T1D shows 15\% higher error than T2D
%     \end{itemize} >> CC-modify
Model performance varies substantially across demographic strata:
\begin{itemize}
\item \textbf{Age-Gender}: Males 18-39 outperform females by 3.2 mg/dL
\item \textbf{Diabetes Type}: T1D shows 15\% higher error than T2D
\end{itemize}

Two-location tracking:

•In .tex files - Use % commented tracking for in-place edits
•In session logs - Show both commented tracking AND clean before/after for clarity

This format ensures: ✓ Every change is explicitly marked ✓ Before/after states are clear ✓ Changes are git-trackable ✓ User can review modifications before approval ✓ Session logs are self-documenting ✓ LaTeX files still compile with tracking in place

Edit Workflow with LaTeX Comment Tracking

Step 1: Read original content

bash

Read(_sections/05_baseline_results.tex, lines 45-52)

Step 2: Make edit with commented tracking

bash

Edit(_sections/05_baseline_results.tex):
  old_string: "\subsection{Subgroup Performance Analysis}"
  new_string: "% ~~\subsection{Subgroup Performance Analysis}~~
%     \subsection{Subgroup Performance and Generalization} >> CC-modify
\subsection{Subgroup Performance and Generalization}"

Step 3: Log in session file (unformatted for readability)

markdown

Change 1: Updated subsection title
~~\subsection{Subgroup Performance Analysis}~~
    \subsection{Subgroup Performance and Generalization} >> CC-modify

Step 4: Verify compilation

bash

pdflatex fairglucose_icml2026.tex
# Comments are ignored, new code compiles ✓

Cleanup Option (post-session): After session is complete and logged, user can optionally run:

bash

# Remove all % commented tracking from .tex files
sed -i '/^% ~~/d' _sections/*.tex
sed -i '/^% >> CC-modify/d' _sections/*.tex

But tracking can safely stay in files indefinitely - it's just comments.

Session Log Template:

markdown

Paper LaTeX Writing - Session Log
==================================

Session Metadata
----------------
Session ID: 003
Date: 2026-01-17
Task: "Reduce Section 5 to 1.5 pages while keeping key results"
Mode: Hybrid (Plan → Approve → Execute)
Author: >> CC (Claude Code)

Current Paper State
-------------------
- Total Pages: 35
- Target Pages: 8
- Venue: ICML 2026
- Sections: 8 main + 3 appendix

---

Analysis Phase
--------------

> JL: "Reduce Section 5 to 1.5 pages while keeping key results"

>> CC: Analyzing Section 5 (Baseline Results)...

Current State:
- File: _sections/05_baseline_results.tex
- Lines: 287
- Estimated Pages: 3.2
- Target: 1.5 pages (~140 lines)
- Reduction Needed: ~50%

Content Breakdown:
1. Overview paragraph (15 lines)
2. Subgroup Performance subsection (85 lines)
   - Verbose description of disparities
   - Redundant pattern descriptions
3. Generalization subsection (72 lines)
   - Detailed OD/ID analysis
   - Overlaps with main text Figure 2 discussion
4. Event Impact subsection (95 lines)
   - Three paragraphs on event effects
   - Some redundancy with Dataset Design section
5. Model Comparison (20 lines)

Reduction Strategy:
- Keep: Overview, Event Impact (key contribution)
- Condense: Subgroup Performance (50% reduction)
- Merge: Generalization into Subgroup Performance
- Cut: Redundant pattern descriptions, overlapping content

Estimated Savings: 147 lines → Target: 140 lines ✓

---

Reduction Plan
--------------

>> CC: Proposed Changes (for approval)

Change 1: Condense Subgroup Performance
- Before: 85 lines (3 paragraphs of detailed description)
- After: 42 lines (1 concise paragraph + bullet points)
- Savings: 43 lines

Change 2: Merge Generalization into Subgroup Performance
- Before: 72 lines (separate subsection)
- After: 25 lines (integrated into one paragraph)
- Savings: 47 lines

Change 3: Tighten Event Impact
- Before: 95 lines (3 verbose paragraphs)
- After: 60 lines (2 concise paragraphs)
- Savings: 35 lines

Change 4: Remove redundant Figure 2 discussion
- Before: 22 lines
- After: 0 lines (already in Section 3)
- Savings: 22 lines

Total Reduction: 147 lines (51% reduction)
Final Estimate: 140 lines (1.5 pages) ✓

> JL: Approved. Proceed.

---

Execution Phase
---------------

>> CC: Executing changes...

[Change 1/4] Condensing Subgroup Performance subsection

File: _sections/05_baseline_results.tex
Lines: 45-130

BEFORE (85 lines):
```latex
\subsection{Subgroup Performance Analysis}

Across all demographic and clinical subgroups, we observe substantial
heterogeneity in model performance. Male patients in the 18-39 age group
consistently demonstrate lower prediction error compared to their female
counterparts in the same age bracket, with rMSE values differing by an
average of 3.2 mg/dL (p < 0.01). This pattern reverses in the 65+ age
group, where female patients exhibit better model performance...

[... 80 more lines of detailed description ...]

AFTER (42 lines):

latex

\subsection{Subgroup Performance and Generalization}

Model performance varies substantially across demographic strata:
- **Age-Gender Interaction**: Males (18-39) outperform females by 3.2 mg/dL
  rMSE; pattern reverses in 65+ group
- **Diabetes Type**: T1D patients show 15% higher error than T2D across all
  models despite structured insulin regimens
- **Generalization**: OD/ID error ratios span 0.8-1.4 across subgroups; T1D
  females (40-64) show greatest degradation under distribution shift

These intersectional disparities highlight the need for fairness-aware
evaluation beyond marginal attributes.

DIFF:

•Lines reduced: 85 → 42 (-43 lines, -51%)
•Content preserved: All key findings retained in bullet format
•Merged: Generalization findings integrated
•Improved: Scannable structure, removed verbose descriptions

✓ Change 1 complete

[Change 2/4] Already integrated into Change 1 ✓ Change 2 complete

[Change 3/4] Tightening Event Impact subsection ...

Session Summary

CC: Session 003 Complete

Changes Made:

•Sections Modified: 1 (_sections/05_baseline_results.tex)
•Lines Changed: 147 removed, 85 added
•Net Reduction: 62 lines
•Commits: 1 (with detailed commit message)

Metrics:

•Section 5 Before: 287 lines (3.2 pages)
•Section 5 After: 140 lines (1.5 pages)
•Target Achievement: ✓ (1.5 pages)
•Overall Paper: 35 pages → 33.3 pages

Next Steps:

•Recompile PDF to verify page count
•Review Section 3 for additional reduction opportunities
•Consider moving detailed results to appendix

Git Status: M _sections/05_baseline_results.tex

Suggested commit message:

code

Reduce Section 5 (Baseline Results) to 1.5 pages

- Condense Subgroup Performance to bullet points
- Merge Generalization into Subgroup Performance
- Tighten Event Impact description
- Remove redundant Figure 2 discussion

Session: 003
Reduction: 287 → 140 lines (51% cut)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

===================================================================

code


## Invocation Patterns

### Natural Language Examples

```bash
# Length reduction
/paper-latex-writing "reduce the paper from 35 to 8 pages"
/paper-latex-writing "shorten Section 3 by 50%"
/paper-latex-writing "cut Introduction to 1 page"

# Clarity improvement
/paper-latex-writing "improve clarity in the Discussion section"
/paper-latex-writing "simplify the methodology description"
/paper-latex-writing "make the abstract more accessible"

# Citation fixes
/paper-latex-writing "fix all citation errors"
/paper-latex-writing "add citations for claims in Section 2"
/paper-latex-writing "ensure all references compile correctly"

# Structural changes
/paper-latex-writing "move the event characterization to Section 3"
/paper-latex-writing "merge subsections 4.1 and 4.2"
/paper-latex-writing "reorganize Related Work by theme"

# Combined tasks
/paper-latex-writing "reduce to 10 pages and fix all citations"
/paper-latex-writing "tighten Section 5, improve clarity, ensure all figures are referenced"

# Global operations
/paper-latex-writing "analyze the entire paper for reduction opportunities"
/paper-latex-writing "check terminology consistency throughout"

Execution Workflow

Hybrid Mode (Plan → Approve → Execute)

code

┌─────────────────────────────────────────────┐
│ 1. PARSE TASK                               │
│    - Extract intent from natural language   │
│    - Identify scope (section/global)        │
│    - Determine primary capability needed    │
└─────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────┐
│ 2. AUTO-DETECT / LOAD CONFIG                │
│    - Find .paper-latex-config.yaml          │
│    - If missing: auto-detect structure      │
│    - Create config for future use           │
└─────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────┐
│ 3. ANALYSIS PHASE                           │
│    - Read relevant sections                 │
│    - Measure current metrics                │
│    - Identify issues/opportunities          │
│    - Log analysis to session file           │
└─────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────┐
│ 4. PLANNING PHASE                           │
│    - Create detailed change plan            │
│    - Estimate impact (lines/pages saved)    │
│    - Show before/after previews             │
│    - Log plan with numbered changes         │
└─────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────┐
│ 5. APPROVAL CHECKPOINT                      │
│    - Present plan to user                   │
│    - Wait for: Approve / Modify / Reject    │
│    - If modifications: return to Planning   │
└─────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────┐
│ 6. EXECUTION PHASE                          │
│    - Execute each change sequentially       │
│    - Log each change with diffs             │
│    - Update metrics after each change       │
│    - Verify compilation after edits         │
└─────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────┐
│ 7. SUMMARY & COMMIT                         │
│    - Summarize all changes                  │
│    - Report final metrics                   │
│    - Generate commit message                │
│    - Archive session log                    │
└─────────────────────────────────────────────┘

Special Features

Length Reduction Strategies

Strategy 1: Analyze First

bash

/paper-latex-writing "analyze paper for reduction to 8 pages"

•Provides breakdown of each section
•Identifies low-value content
•Suggests specific cuts
•No changes made (analysis only)

Strategy 2: Global Plan

bash

/paper-latex-writing "create reduction plan to reach 8 pages"

•Multi-section coordinated plan
•Balanced reduction across sections
•Preserves paper flow
•Requires approval before execution

Strategy 3: Iterative Reduction

bash

/paper-latex-writing "reduce paper by 10%"
# Review results
/paper-latex-writing "reduce by another 10%"
# Repeat until target reached

•Gentle, controlled reduction
•Easy to review incremental changes
•Can stop when quality starts degrading

Metrics Tracking

Automatically tracked across sessions:

yaml

metrics:
  total_pages: 35 → 33.3 → 30.1 → 28.5
  total_lines: 4200 → 3980 → 3654 → 3421
  sections_modified: [5, 3, 1, 2]
  citations_fixed: 12
  figures_checked: 18
  compilation_success: true
  sessions_completed: 4

Diff Visualization

Each change shows:

diff

File: _sections/03_dataset_design.tex
Lines: 45-68

- Despite this advancement, existing CGM modeling benchmarks are hindered by
- unbalanced datasets, limited demographic diversity, and lack of event
- annotations (meals, exercise, medication), obscuring both performance
- disparities across patient types and prediction challenges driven by
- glucose-altering events.
+ Existing CGM benchmarks lack demographic balance and event annotations,
+ obscuring both fairness issues and event-driven prediction challenges.

Lines: 2 → 1 (-1 line, -50%)
Words: 48 → 16 (-32 words, -67%)

Integration with Git

Commit Message Generation

After each session, generates detailed commit message:

code

Reduce Section 5 and fix citations (Session 003)

Length Reduction:
- Section 5: 287 → 140 lines (1.5 pages)
- Condensed Subgroup Performance to bullets
- Merged Generalization subsection
- Removed redundant Figure 2 discussion

Citation Fixes:
- Fixed 3 broken citations (tft, informer)
- Updated bibliography formatting
- Added missing citations in Related Work

Metrics:
- Total pages: 35 → 33.3 (-5%)
- Compilation: ✓ successful
- Session log: paper/code-logging/archives/session-003-length-reduction.md

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Configuration Reference

`.paper-latex-config.yaml` Schema

yaml

# Paper identification
main_tex: string              # Main .tex file
bibliography: string          # .bib file
structure: modular|monolithic # File organization

# Directories
sections_dir: string          # Modular sections folder
figures_dir: string
tables_dir: string
appendix_figures_dir: string
appendix_tables_dir: string
logging_dir: string           # Session logs location

# Section mapping (for modular papers)
sections:
  section_name: path/to/file.tex

appendix_sections:
  appendix_name: path/to/file.tex

# Compilation
compiler: pdflatex|xelatex|lualatex
bibtex: bibtex|biber
compile_flags: [string]

# Target metrics
target_pages: integer
current_pages: integer
venue: string
track: string

# Style preferences (optional)
style:
  max_sentence_length: integer
  max_paragraph_lines: integer
  prefer_active_voice: boolean
  terminology_map:
    dataset: "dataset"  # Enforce consistency
    data-set: "dataset"

# Custom ignore patterns
ignore_sections:
  - acknowledgments
  - references

Best Practices

1. Start with Analysis

Always run analysis before making changes:

bash

/paper-latex-writing "analyze paper for 8-page reduction"
# Review analysis
/paper-latex-writing "execute the reduction plan"

2. One Capability at a Time

For complex tasks, separate concerns:

bash

# Bad: Too many goals
/paper-latex-writing "reduce length, fix all citations, reorganize structure"

# Good: Sequential focused tasks
/paper-latex-writing "reduce to target length"
/paper-latex-writing "fix all citations"
/paper-latex-writing "reorganize Related Work"

3. Verify After Each Session

Always compile and review:

bash

pdflatex fairglucose_icml2026.tex
# Check page count, references, formatting

4. Use Git for Safety

Commit before major operations:

bash

git add .
git commit -m "Before length reduction session"
/paper-latex-writing "reduce by 30%"
# If unhappy with results: git reset --hard HEAD

5. Iterative for Large Cuts

For large reductions (35 → 8 pages), iterate:

bash

/paper-latex-writing "reduce to 20 pages"  # -43%
# Review and adjust
/paper-latex-writing "reduce to 12 pages"  # -40%
# Review and adjust
/paper-latex-writing "reduce to 8 pages"   # -33%

Error Handling

Common Issues

Issue: Main .tex file not found

•Detection: Auto-scan finds multiple .tex files or none
•Resolution: Prompt user to specify, create config

Issue: Compilation fails after edits

•Detection: Run pdflatex after each change
•Resolution: Automatic rollback of last change, log error

Issue: Section file missing

•Detection: \input{} references non-existent file
•Resolution: Update config, skip missing sections

Issue: Citation key not found in .bib

•Detection: Parse .bib and cross-check \cite{} commands
•Resolution: List missing keys, ask user to add or remove citations

Advanced Usage

Custom Reduction Rules

Create .paper-latex-rules.yaml:

yaml

reduction_priorities:
  high:
    - verbose_descriptions
    - redundant_citations
    - overlapping_content
  medium:
    - example_details
    - background_material
  low:
    - key_contributions
    - novel_results
    - fairness_findings

protected_content:
  - \section{Introduction}   # Never auto-cut introduction
  - Figure 1                  # Key figure must stay
  - Table 2                   # Main results table

cut_candidates:
  - "Moreover,"      # Remove filler transitions
  - "It is worth noting that"
  - "In other words"

Batch Operations

Process multiple papers:

bash

/paper-latex-writing --config paper1-config.yaml "reduce to 8 pages"
/paper-latex-writing --config paper2-config.yaml "reduce to 8 pages"

Tool Requirements

The skill should use these Claude Code tools:

•Read: Read .tex, .bib, config files
•Edit: Make precise text replacements
•Write: Create config files, session logs
•Bash: Run pdflatex, bibtex, check compilation
•Glob: Find all .tex files, figures
•Grep: Search for citations, references, patterns

Output Format

Always respond with:

•Session header with task and metadata
•Analysis summary with current state
•Proposed plan with estimated impact
•Approval request (if hybrid mode)
•Execution log with diffs
•Final summary with metrics and commit message

Implementation Notes

This skill should be invoked via:

bash

/paper-latex-writing "natural language task description"

The skill activates when user provides paper editing task in natural language. All changes are logged, trackable via git, and reversible.

paper-latex-writing

Purpose

Capabilities

1. Length Reduction

2. Clarity Improvement

3. Citation Management

4. Structural Reorganization

5. Figure/Table Management

6. Consistency Checking

Paper Structure Auto-Detection

Hybrid Detection Strategy

Fallback Behavior

Session-Based Logging

Log File Structure

Log Format

Edit Workflow with LaTeX Comment Tracking

Session Summary

Execution Workflow

Hybrid Mode (Plan → Approve → Execute)

Special Features

Length Reduction Strategies

Metrics Tracking

Diff Visualization

Integration with Git

Commit Message Generation

Configuration Reference

.paper-latex-config.yaml Schema

Best Practices

1. Start with Analysis

2. One Capability at a Time

3. Verify After Each Session

4. Use Git for Safety

5. Iterative for Large Cuts

Error Handling

Common Issues

Advanced Usage

Custom Reduction Rules

Batch Operations

Tool Requirements

Output Format

Implementation Notes

`.paper-latex-config.yaml` Schema