Paper Write

Override for Codex users who want Gemini, not a second Codex agent, to act as the reviewer. Install this package after skills/skills-codex/*.

Paper Write: Section-by-Section LaTeX Generation

Gemini overlay assurance: review_independence: cross-family and acceptance_status: accepted.

Draft a LaTeX paper based on: $ARGUMENTS

Constants

• REVIEWER_MODEL = gemini-review — Gemini reviewer invoked through the local gemini-review MCP bridge. Set GEMINI_REVIEW_MODEL if you need a specific Gemini model override.
• TARGET_VENUE = ICLR — Default venue. Supported: ICLR, NeurIPS, ICML. Determines style file and formatting.
• ANONYMOUS = true — If true, use anonymous author block. Set false for camera-ready.
• MAX_PAGES = 9 — Main body page limit. Counts from first page to end of Conclusion section. References and appendix are NOT counted.
• DBLP_BIBTEX = true — Fetch real BibTeX from DBLP/CrossRef instead of LLM-generated entries. Eliminates hallucinated citations. Zero install required. Set false to use legacy behavior (LLM search + [VERIFY] markers).

Inputs

1. PAPER_PLAN.md — outline with claims-evidence matrix, section plan, figure plan (from /paper-plan)
2. NARRATIVE_REPORT.md — the research narrative (primary source of content)
3. Generated figures — PDF/PNG files in figures/ (from /paper-figure)
4. LaTeX includes — figures/latex_includes.tex (from /paper-figure)
5. Bibliography — existing .bib file, or will create one

If no PAPER_PLAN.md exists, ask the user to run /paper-plan first or provide a brief outline.

Templates

Venue-Specific Setup

The skill includes conference templates in templates/. Select based on TARGET_VENUE:

ICLR:

\documentclass{article}
\usepackage{iclr2026_conference,times}
% \iclrfinalcopy  % Uncomment for camera-ready

NeurIPS:

\documentclass{article}
\usepackage[preprint]{neurips_2025}
% \usepackage[final]{neurips_2025}  % Camera-ready

ICML:

\documentclass[accepted]{icml2025}
% Use [accepted] for camera-ready

Project Structure

Generate this file structure:

paper/
├── main.tex                    # master file (includes sections)
├── iclr2026_conference.sty     # or neurips_2025.sty / icml2025.sty
├── math_commands.tex           # shared math macros
├── references.bib              # bibliography (filtered — only cited entries)
├── sections/
│   ├── 0_abstract.tex
│   ├── 1_introduction.tex
│   ├── 2_related_work.tex
│   ├── 3_method.tex            # or preliminaries, setup, etc.
│   ├── 4_experiments.tex
│   ├── 5_conclusion.tex
│   └── A_appendix.tex          # proof details, extra experiments
└── figures/                    # symlink or copy from project figures/

Section files are FLEXIBLE: If the paper plan has 6-8 sections, create corresponding files (e.g., 4_theory.tex, 5_experiments.tex, 6_analysis.tex, 7_conclusion.tex).

Workflow

Step 0: Backup and Clean

If paper/ already exists, back up to paper-backup-{timestamp}/ before overwriting. Never silently destroy existing work.

CRITICAL: Clean stale files. When changing section structure (e.g., 5 sections → 7 sections), delete section files that are no longer referenced by main.tex. Stale files (e.g., old 5_conclusion.tex left behind when conclusion moved to 7_conclusion.tex) cause confusion and waste space.

Step 1: Initialize Project

1. Create paper/ directory
2. Copy venue template from templates/ — the template already includes:
   • All standard packages (amsmath, hyperref, cleveref, booktabs, etc.)
   • Theorem environments with \crefname{assumption} fix
   • Anonymous author block
3. Generate math_commands.tex with paper-specific notation
4. Create section files matching PAPER_PLAN structure

Author block (anonymous mode):

\author{Anonymous Authors}

Step 2: Generate math_commands.tex

Create shared math macros based on the paper's notation:

% math_commands.tex — shared notation
\newcommand{\R}{\mathbb{R}}
\newcommand{\E}{\mathbb{E}}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\argmax}{arg\,max}
% Add paper-specific notation here

Step 3: Write Each Section

Process sections in order. For each section:

1. Read the plan — what claims, evidence, citations belong here
2. Read NARRATIVE_REPORT.md — extract relevant content, findings, and quantitative results
3. Draft content — write complete LaTeX (not placeholders)
4. Insert figures/tables — use snippets from figures/latex_includes.tex
5. Add citations — use \citep{} / \citet{} (all three venues use natbib)

Section-Specific Guidelines

§0 Abstract:

• Must be self-contained (understandable without reading the paper)
• Structure: problem → approach → key result → implication
• Include one concrete quantitative result
• 150-250 words (check venue limit)
• No citations, no undefined acronyms
• No \begin{abstract} — that's in main.tex

§1 Introduction:

• Open with a compelling hook (1-2 sentences, problem motivation)
• State the gap clearly ("However, ...")
• List contributions as a numbered or bulleted list
• End with a brief roadmap ("The rest of this paper is organized as...")
• Include the main result figure if space allows
• Target: 1.5 pages

§2 Related Work:

• MINIMUM 1 full page (3-4 substantive paragraphs). Short related work sections are a common reviewer complaint.
• Organize by category using \paragraph{Category Name.}
• Each category: 1 paragraph summarizing the line of work + 1-2 sentences positioning this paper
• Do NOT just list papers — synthesize and compare
• End each paragraph with how this paper relates/differs

§3 Method / Preliminaries / Setup:

• Define notation early (reference math_commands.tex)
• Use \begin{definition}, \begin{theorem} environments for formal statements
• For theory papers: include proof sketches of key results in main body, full proofs in appendix
• For theory papers: include a comparison table of prior bounds vs. this paper
• Include algorithm pseudocode if applicable (algorithm2e or algorithmic)
• Target: 1.5-2 pages

§4 Experiments:

• Start with experimental setup (datasets, baselines, metrics, implementation details)
• Main results table/figure first
• Then ablations and analysis
• Every claim from the introduction must have supporting evidence here
• Target: 2.5-3 pages

§5 Conclusion:

• Summarize contributions (NOT copy-paste from intro — rephrase)
• Limitations (2-4 material, specific limits — dataset scale, compute regime, assumption X; real ones only, never invented to fill a count. This section is the ONLY home for generic caveats)
• Future work (1-2 concrete directions)
• Ethics statement and reproducibility statement (if venue requires)
• Target: 0.5 pages

Appendix:

• Proof details (full proofs of main-body theorems)
• Additional experiments, ablations
• Implementation details, hyperparameter tables
• Additional visualizations

Step 4: Build Bibliography

CRITICAL: Only include entries that are actually cited in the paper.

1. Scan all \citep{} and \citet{} references in the drafted sections
2. Build a citation key list
3. For each citation key:
   • Check existing .bib files in the project/narrative docs
   • If not found and DBLP_BIBTEX = true, use the verified fetch chain below
   • If not found and DBLP_BIBTEX = false, search arXiv/Scholar for correct BibTeX
   • NEVER fabricate BibTeX entries — mark unknown ones with [VERIFY] comment
4. Write references.bib containing ONLY cited entries (no bloat)

Verified BibTeX Fetch (when DBLP_BIBTEX = true)

Three-step fallback chain — zero install, zero auth, all real BibTeX:

Step A: DBLP (best quality — full venue, pages, editors)

# 1. Search by title + first author
curl -s "https://dblp.org/search/publ/api?q=TITLE+AUTHOR&format=json&h=3"
# 2. Extract DBLP key from result (e.g., conf/nips/VaswaniSPUJGKP17)
# 3. Fetch real BibTeX
curl -s "https://dblp.org/rec/{key}.bib"

Step B: CrossRef DOI (fallback — works for arXiv preprints)

# If paper has a DOI or arXiv ID (arXiv DOI = 10.48550/arXiv.{id})
curl -sLH "Accept: application/x-bibtex" "https://doi.org/{doi}"

Step C: Mark [VERIFY] (last resort) If both DBLP and CrossRef return nothing, mark the entry with % [VERIFY] comment. Do NOT fabricate.

Why this matters: LLM-generated BibTeX frequently hallucinates venue names, page numbers, or even co-authors. DBLP and CrossRef return publisher-verified metadata. Upstream skills (/research-lit, /novelty-check) may mention papers from LLM memory — this fetch chain is the gate that prevents hallucinated citations from entering the final .bib.

Automated bib cleaning — use this Python pattern to extract only cited entries:

import re
# 1. Grep all \citep{...} and \citet{...} from all .tex files
# 2. Extract unique keys (handle multi-cite like \citep{a,b,c})
# 3. Parse the full .bib file, keep only entries whose key is in the cited set
# 4. Write the filtered bib

This prevents bib bloat (e.g., 948 lines → 215 lines in testing).

Citation verification rules (from claude-scholar + Imbad0202):

1. Every BibTeX entry must have: author, title, year, venue/journal
2. Prefer published venue versions over arXiv preprints (if published)
3. Use consistent key format: {firstauthor}{year}{keyword} (e.g., ho2020denoising)
4. Double-check year and venue for every entry
5. Remove duplicate entries (same paper with different keys)

Step 5: De-AI Polish (from kgraph57/paper-writer-skill)

After drafting all sections, scan for common AI writing patterns and fix them:

Content patterns to fix:

• Significance inflation ("groundbreaking", "revolutionary" → use measured language)
• Formulaic transitions ("In this section, we..." → remove or vary)
• Generic conclusions ("This work opens exciting new avenues" → be specific)

Language patterns to fix (watch words):

• Replace: delve, pivotal, landscape, tapestry, underscore, noteworthy, intriguingly
• Remove filler: "It is worth noting that", "Importantly,", "Notably,"
• Avoid rule-of-three lists ("X, Y, and Z" appearing repeatedly)
• Don't start consecutive sentences with "This" or "We"

Step 6: Cross-Review with REVIEWER_MODEL

Send the complete draft to Gemini review:

mcp__gemini-review__review_start:
  prompt: |
    Review this [VENUE] paper draft (main body, excluding appendix).

    Judge claim calibration in BOTH directions. Recommend narrowing only when the
    current scope or modality exceeds the evidence; do not ask for extra hedges
    around a supported result. Flag stacked hedges, self-defence ("we do not
    claim"), instruction confessions ("we do not address X"), and generic caveats
    outside Limitations as writing defects to remove. Tone fixes must never alter
    facts, negation, modality, scope, comparison direction, or numbers.

    Focus on:
    1. Does each claim from the intro have supporting evidence?
    2. Is the writing clear, concise, and free of AI-isms?
    3. Any logical gaps or unclear explanations?
    4. Does it fit within [MAX_PAGES] pages (to end of Conclusion)?
    5. Is related work sufficiently comprehensive (≥1 page)?
    6. For theory papers: are proof sketches adequate?
    7. Are figures/tables clearly described and properly referenced?

    For each issue, specify: severity (CRITICAL/MAJOR/MINOR), location, and fix.

    [paste full draft text]

After this start call, immediately save the returned jobId and poll mcp__gemini-review__review_status with a bounded waitSeconds until done=true. Treat the completed status payload's response as the reviewer output, and save the completed threadId for any follow-up round.

Apply CRITICAL and MAJOR fixes. Document MINOR issues for the user.

Step 7: Reverse Outline Test (from Research-Paper-Writing-Skills)

After drafting all sections:

1. Extract topic sentences — pull the first sentence of every paragraph
2. Read them in sequence — they should form a coherent narrative on their own
3. Check claim coverage — every claim from the Claims-Evidence Matrix must appear
4. Check evidence mapping — every experiment/figure must support a stated claim
5. Fix gaps — if a topic sentence doesn't advance the story, rewrite the paragraph

Step 8: Final Checks

Before declaring done:

• All \ref{} and \label{} match (no undefined references)
• All \citep{} / \citet{} have corresponding BibTeX entries
• No author information in anonymous mode
• Figure/table numbering is correct
• Page count within MAX_PAGES (main body to Conclusion end)
• No TODO/FIXME/XXX markers left in the text
• No [VERIFY] markers left unchecked
• Abstract is self-contained (understandable without reading the paper)
• Title is specific and informative (not generic)
• Related work is ≥1 full page
• references.bib contains ONLY cited entries (no bloat)
• No stale section files — every .tex in sections/ is \inputed by main.tex
• Section files match main.tex — file numbering and \input paths are consistent

Output Protocols

Follow these shared protocols for all output files:

• Output Versioning Protocol — write timestamped file first, then copy to fixed name
• Output Manifest Protocol — log every output to MANIFEST.md
• Output Language Protocol — respect the project's language setting

Key Rules

=== CONFIDENT PROSE, HONEST LIMITS (never upgrades claims) ===

1. Calibrate each claim to the evidence's actual scope and modality, then state that calibrated claim directly. Necessary assumptions, uncertainty, and scope are part of the claim; stacked hedges and defensive throat-clearing are not.
2. If the current claim is unsupported, narrow it to a version the evidence supports or cut it. Do not substitute a softer-sounding synonym for fixing scope, modality, comparison, or aggregation.
3. Put generic caveats and broader boundary discussion in one Limitations section. Outside it, remove generic disclaimers such as "further research is needed", "may not generalize", and "should be interpreted with caution". Claim-defining scope, assumptions, and statistical qualifications stay attached to the claims they make true.
4. Aim for 2-4 material, specific limitations (dataset scale, compute regime, assumption X). Real ones only — never invent one to meet a count, never apologize generically, never repeat the same limitation through the paper.
5. Writing instructions are not manuscript content. "Do not mention X" means omit X, not write "we do not address/claim/discuss X". Never expose drafting instructions, requested omissions, reviewer feedback, or revision history in manuscript prose.
6. Replace self-defence ("we do not claim", "our goal is merely") with a positive, evidence-matched statement of what the paper does establish. If the defensive sentence carries a real boundary, keep that boundary in the claim or Limitations; do not delete truth-conditional content.
7. Tone-only edits never alter facts, negation, modality, scope, assumptions, comparison direction, aggregation, numbers, formulas, or citations. Genuine overclaims must still be narrowed; supported claims wrapped in redundant caution must be stated directly.
8. One causal spine: gap -> question -> insight -> consequence -> evidence -> implication. Every section advances it. Make the method feel inevitable: the gap creates a concrete question, the key insight answers it, the method follows from the insight, each major experiment tests a consequence of it, and the conclusion states exactly what the evidence establishes. Front-load the contribution; never narrate the drafting or revision process.
9. The paper is a launch, not a progress report. Organize the narrative around the work's strongest genuine advantage — a new capability, problem, mechanism or viewpoint, wider applicability, lower cost, a better tradeoff. Material that does not form an advantage stays out of the main line. If the results cannot carry the original story, rebuild the story around the strongest evidence instead of defending the original one.
10. Pick the contest the paper wins. Do not build the narrative on a metric where the method is not ahead; frame the comparison around the task definition, evaluation dimension or constraint that reflects what the method is for, and say explicitly which contest it wins. Unfavorable numbers still appear — tables stay complete. Where the evidence supports it, explain them as a goal difference or a deliberate tradeoff rather than narrating a defeat ("underperforms", "fails to surpass"); where it does not, state the underperformance neutrally, narrow the claim, and keep it in Limitations if it is material. Never elevate a local observation into a verdict on the whole method, and never invent a tradeoff to cover a weakness.
11. Every experiment has an argumentative duty: it shows the method works, shows the gain comes from the key mechanism, shows value in the target scenario, or rules out the most likely alternative explanation. An experiment carrying none of these is cut, shortened, moved to the appendix, or redesigned. The experiments section is an argument, not a results warehouse.
12. State the advantage yourself — under which condition it appears, why it appears, what it solves — rather than expecting the reviewer to find it in a table. Abstract and introduction open like a launch: an important unsolved problem, the gap in existing methods, this paper's distinct idea, the heaviest result. The conclusion reinforces what was solved, proposed and proven and why it matters; no new self-negation or widened limitations in the last paragraph.

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.

• Do NOT generate author names, emails, or affiliations — use anonymous block or placeholder

• Write complete sections, not outlines — the output should be compilable LaTeX

• One file per section — modular structure for easy editing

• Every claim must cite evidence — cross-reference the Claims-Evidence Matrix

• Compile-ready — the output should compile with latexmk without errors (modulo missing figures)

• Calibrate, don't hedge — match each claim to its evidence's actual scope and modality, then state it directly; generic caveats live in Limitations only (the CONFIDENT PROSE, HONEST LIMITS block above is the contract)

• Launch, not progress report — organize around the strongest genuine advantage, pick the contest the paper wins, give every experiment an argumentative duty; unfavorable numbers stay in the tables, explained as tradeoffs where the evidence supports that and stated neutrally where it does not — never narrated as defeats, never dressed as a tradeoff they are not (rules 9-12 above)

• Order results by argument, not by lab notebook — present experiments in the sequence that best builds the case, never in the order they happened to run

• Controls and ablations sit next to the claim they test — not pooled in a distant subsection where the reader has forgotten what was at stake

• Venue style matters — all three venues (ICLR/NeurIPS/ICML) use natbib (\citep/\citet)

• Page limit = main body to Conclusion — references and appendix do NOT count

• Clean bib — references.bib must only contain entries that are actually \cited

• Section count is flexible — match PAPER_PLAN structure, don't force into 5 sections

• Backup before overwrite — never destroy existing paper/ directory without backing up

Writing Quality Reference

Principles from Research-Paper-Writing-Skills:

1. One message per paragraph — each paragraph makes exactly one point
2. Topic sentence first — the first sentence states the paragraph's message
3. Explicit transitions — connect paragraphs with logical connectors
4. Reverse outline test — extract topic sentences; they should form a coherent narrative

De-AI patterns from kgraph57/paper-writer-skill:

5. No AI watch words — delve, pivotal, landscape, tapestry, underscore
6. No significance inflation — groundbreaking, revolutionary, paradigm shift
7. No formulaic structures — vary sentence openings and transitions

Acknowledgements

Writing methodology adapted from Research-Paper-Writing-Skills (CCF award-winning methodology). Citation verification from claude-scholar and Imbad0202/academic-research-skills. De-AI polish from kgraph57/paper-writer-skill. Backup mechanism from baoyu-skills.