Paper Write Agent Skill

Paper Write: academic paper writing and LaTeX setup

For LaTeX environment setup, biber vs bibtex, venue templates and compilation, follow references/latex.md.

Core Principles

• Precision over elegance. Every claim traces to code, data or a cited source.
• Conciseness. Remove words that don't add information.

Writing Process

• Read first. Always read the existing section before writing.
• Identify the claim. What is this paragraph trying to say?
• Find the evidence. Where in code/results does this come from?
• Write the text. State claim, present evidence, interpret.
• Verify. Re-read against source to catch any drift, then compile with the project pipeline from references/latex.md Step 5 and fix any error before the next edit.

Section-Specific Guidance

Abstract

• Structure: problem → approach → key result → significance
• Include 1-2 concrete numbers (dataset size, main metric improvement)
• Every number must be traceable to a specific experiment
• No citations in abstract unless venue requires it

Introduction

• Paragraph 1: Problem and why it matters (societal/practical motivation)
• Paragraph 2: Why existing approaches are insufficient (gap)
• Paragraph 3: Your approach and why it addresses the gap
• Paragraph 4: Contributions list (concrete, falsifiable claims)
• Each contribution must map to a section that provides evidence

Related Work

• Organize by theme/approach, not chronologically
• For each group: what they do, what's missing, how your work differs
• Be fair: acknowledge strengths of prior work, don't strawman
• End each paragraph with how your work addresses the limitation

Methods

• Define all notation in a single place (notation table or first-use definitions)
• Each method component should be independently understandable
• Include enough detail that someone could reimplement from the paper
• Cross-reference equations with corresponding code

Experiments

• Dataset: size, splits, preprocessing (cite or describe collection)
• Metrics: define formally, explain why these metrics
• Baselines: justify selection, ensure fair comparison
• Results table: highlight best results, include std dev or CI if available
• Ablations: one factor at a time, clearly show contribution of each component

Conclusion

• Summarize contributions (not the entire paper)
• State limitations honestly
• Future work: specific and feasible, not vague

Notation Consistency Protocol

Define once, use identically everywhere. When writing or editing any section:

• Read existing notation definitions in the paper
• Use EXACTLY the same symbols. Do not introduce synonyms.
• If a new symbol is needed, check it doesn't clash with existing ones
• Maintain a notation table if the paper has one

Common pitfalls:

• Using both $x$ and $\mathbf{x}$ for the same concept
• Defining $N$ as dataset size in methods but using $n$ in experiments
• Inconsistent subscript conventions (e.g., $f_i$ vs $f(i)$)

Figure Refinement Methodology

1. Specification Capture

Before generating or modifying any figure:

• What data does it show? (exact source file/variable)
• What message should the reader take away?
• What are the hard constraints? (font size ≥ 8pt, column width, color scheme)
• What aspects of the current version are correct and must be preserved?

2. Constraint Preservation

Across multiple rounds of revision, track constraints explicitly:

Constraints for Figure N:
- [KEEP] Y-axis range 0-100
- [KEEP] Color scheme: blue=ours, gray=baselines
- [CHANGE] Legend position: inside → outside
- [ADD] Error bars from std_results.json

3. Variant Generation

When exploring design alternatives:

• Generate 2-3 variants side by side when feasible
• Each variant changes ONE visual aspect
• Let the user compare and choose, don't pick for them

4. Visual Verification

After generating any figure, ALWAYS:

• Open the rendered image and look at it
• Check data values, colors and axis ranges against the spec and constraint list above
• Verify labels, legends, and annotations are correct
• Confirm the takeaway message is clear from a glance

Output Format

When writing paper text:

• Provide LaTeX-ready output that matches the paper's existing style
• Include comments for any claim that needs verification: % TODO: verify this number
• Flag any notation inconsistencies found during writing
• Suggest specific improvements with before/after comparisons