paper-write

name	paper-write
description	Draft LaTeX paper section by section from an outline. Use when user says \"写论文\", \"write paper\", \"draft LaTeX\", \"开始写\", or wants to generate LaTeX content from a paper plan.

Paper Write: Section-by-Section LaTeX Generation

Draft a LaTeX paper based on: $ARGUMENTS

Constants

• REVIEWER_MODEL = gpt-5.5 — Model used via a secondary Codex agent for section review. Must be an OpenAI model.
• TARGET_VENUE = ICLR — Default venue. Supported: ICLR, NeurIPS, ICML, CVPR (also ICCV/ECCV), ACL (also EMNLP/NAACL), AAAI, ACM (ACM MM, SIGIR, KDD, CHI, etc.), IEEE_JOURNAL (IEEE Transactions / Letters, e.g., T-PAMI, JSAC, TWC, TCOM, TSP, TIP), IEEE_CONF (IEEE conferences, e.g., ICC, GLOBECOM, INFOCOM, ICASSP). Determines style file and formatting.
• ANONYMOUS = true — If true, use anonymous author block. Set false for camera-ready. Note: most IEEE venues do NOT use anonymous submission — set false for IEEE.
• MAX_PAGES = 9 — Main body page limit. For ML conferences: counts from first page to end of Conclusion section, references and appendix NOT counted. For IEEE venues: references ARE counted toward the page limit. Typical limits: IEEE journal = no strict limit (but 12-14 pages typical for Transactions, 4-5 for Letters), IEEE conference = 5-8 pages including references.
• DBLP_BIBTEX = true — Fetch real BibTeX from DBLP/CrossRef instead of LLM-generated entries. Eliminates hallucinated citations. Zero install required. Final bibliography entries default to published conference/journal versions; arXiv is a discovery source only unless the user explicitly accepts an unpublished-preprint exception.

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.

Draft Readiness and Composition Order

Before prose, re-read the one-sentence contribution and the Claim–Evidence–Figure Storyboard in PAPER_PLAN.md. If the research question, main finding, significance, figure evidence, or permitted conclusion is still unstable, stop and return to /paper-plan or the evidence audit; do not write around an unresolved result.

Compose the first draft in evidence order, while preserving the plan's conventional final section order in main.tex:

1. frozen contribution and figure storyboard;
2. Methods / setup needed to interpret the evidence;
3. Results in storyboard order;
4. Discussion (interpretation, literature comparison, limitations);
5. Introduction;
6. Abstract and title.

This is a composition sequence, not a requirement to reorder the submitted paper. It prevents an attractive Introduction from getting ahead of the available evidence.

Orchestra-Guided Writing Overlay

Keep the existing workflow, file layout, and defaults. Use the shared references below only when they improve writing quality:

• Read ../../skills-codex/shared-references/writing-principles.md before drafting the Abstract, Introduction, Related Work, or when prose feels generic
• Read ../../skills-codex/shared-references/paper-writing-rules.md before drafting section prose, citations, math notation, figures/tables, and experiment discussion
• Read ../../skills-codex/shared-references/icde-yu-memory-paper-structure.md before initializing a new IEEE/Overleaf paper or drafting a systems, memory, query-processing, or benchmark paper
• Read ../../skills-codex/shared-references/venue-checklists.md during the final write-up and submission-readiness pass
• Read ../../skills-codex/shared-references/citation-discipline.md only when the built-in DBLP/CrossRef workflow is insufficient

These references are support material, not extra workflow phases.

Templates

Venue-Specific Setup

The skill includes modular paper-package templates under templates/. Select the concrete skeleton based on TARGET_VENUE:

• IEEE_CONF -> templates/ieee_conference.tex
• IEEE_JOURNAL -> templates/ieee_journal.tex
• non-IEEE venues may use the venue's official template, but should still follow the same modular package conventions: thin main.tex, section files, dedicated figure/table assets, and appendix material isolated from the main narrative

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

IEEE Journal (Transactions, Letters):

\documentclass[journal]{IEEEtran}
\usepackage{cite}  % IEEE uses \cite{}, NOT natbib
% Author block uses \author{Name~\IEEEmembership{Member,~IEEE}}

IEEE Conference (ICC, GLOBECOM, INFOCOM, ICASSP, etc.):

\documentclass[conference]{IEEEtran}
\usepackage{cite}  % IEEE uses \cite{}, NOT natbib
% Author block uses \IEEEauthorblockN / \IEEEauthorblockA

Project Structure

For new IEEE/Overleaf starts, generate this ICDE_YU_Memory-compatible structure unless the target venue package requires different names:

paper/
  main.tex
  IEEEtran.cls
  IEEE.bib
  Content/
    0.Abstract.tex
    1.Introduction.tex
    2.Methodology.tex
    3.BenchmarkConstruction.tex
    4.Experiment.tex
    5.Discussion.tex
    6.RelatedWork.tex
    7.Conclusion.tex
  Figure/
    figure1.pdf
    figure2.pdf

For non-IEEE or existing pipeline projects, this normalized structure remains acceptable:

paper/
├── main.tex                    # master file (includes sections)
├── iclr2026_conference.sty     # or neurips_2025.sty / icml2025.sty / IEEEtran.cls + IEEEtran.bst
├── 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).

For IEEE-oriented modular starts, prefer copying from:

• templates/ieee_conference.tex
• templates/ieee_journal.tex

Treat them as package skeletons, then swap in the target venue's official class/style bundle if submission instructions require a stricter venue package.

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 plus Content/ and Figure/ for new IEEE/Overleaf starts; otherwise follow the existing project convention.
2. Copy the venue template from templates/ (or the target venue's official bundle while preserving the same modular layout) — 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 and the selected layout

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

Compose in the evidence order defined above: Methods → Results in storyboard order → Discussion → Introduction → Abstract/title. Then place completed prose in the final section order specified by PAPER_PLAN.md and main.tex.

For each drafting unit:

1. Read the plan and storyboard — identify the supported claim, primary question, evidence locator, permitted conclusion, and excluded conclusion.
2. Read raw-linked evidence — use NARRATIVE_REPORT.md plus the cited run, table, figure source, or audit record; do not rely on a narrative claim when the raw locator conflicts.
3. Draft content — write complete LaTeX (no fabricated placeholders).
4. Insert figures/tables — use snippets from figures/latex_includes.tex. Captions must state the comparison, n/statistics where applicable, and the conclusion readers may take from the figure.
5. Add citations — for ML conferences (ICLR/NeurIPS/ICML/CVPR/ACL/AAAI): use \citep{} / \citet{} (natbib). For IEEE venues: use \cite{} (numeric style via cite package). Never mix natbib and cite commands.

When using the ICDE-style layout, translate figure snippets to Figure/ paths inside \includegraphics and write bibliography entries to IEEE.bib.

Before drafting the front matter, re-read the one-sentence contribution from PAPER_PLAN.md. The Abstract and Introduction should make that takeaway obvious before the reader reaches the full method.

Section-Specific Guidelines

Preferred package standard, matching ICDE_YU_Memory quality or better:

• main.tex stays thin and declarative.
• For new IEEE/Overleaf papers, section prose lives under paper/Content/, figures under paper/Figure/, and bibliography in paper/IEEE.bib.
• Preserve paper/sections/, paper/figures/, and references.bib for existing projects and non-IEEE templates that already use them.
• For system/memory papers, allow names like 2.Methodology.tex, 3.BenchmarkConstruction.tex, 4.Experiment.tex, 5.Discussion.tex, 6.RelatedWork.tex, 7.Conclusion.tex.

For system or benchmark papers, placing Related Work after Experiments is allowed when the technical story reads more cleanly that way.

For system-, memory-, or query-processing papers, write Methodology with explicit substructure instead of a single undifferentiated block. Create a dedicated Benchmark / Dataset Construction section when the benchmark, stress test, evaluation split, or representative-selection protocol is part of the contribution. Structure Experiments by research questions (RQ1, RQ2, ...) when possible, and use a dedicated Discussion section when limitations, trade-offs, or scope boundaries deserve independent treatment.

§0 Abstract:

• Use the 5-part flow from ../../skills-codex/shared-references/writing-principles.md: what, why hard, how, evidence, strongest result
• 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:

• Default compact length: write 2-4 tight synthesis paragraphs. In IEEE/Overleaf-style two-column papers, Related Work plus Conclusion should not exceed about 0.8 page unless the venue or user explicitly asks for a longer survey.
• 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 / Results:

• Start with experimental setup (datasets, baselines, metrics, implementation details)
• Present main results in Claim–Evidence–Figure Storyboard order, then ablations and analysis; structure by research questions (RQ1, RQ2, ...) when possible.
• Every claim from the Introduction must have supporting evidence here.
• For each primary figure/table, write a four-sentence Results paragraph:
  1. state the question; 2. state the comparison or analysis; 3. report the key observation; 4. state only the storyboard's bounded conclusion.
• Results report observations and comparisons. Do not move mechanism explanations, literature comparison, limitations, or causal claims not identified by the storyboard into Results.
• Target: 2.5-3 pages.

Discussion (when planned as a separate section):

• Interpret the observed results, compare with prior work, explain supported differences, and state concrete limitations. Do not repeat the Results or add new unreported evidence.

§5 Conclusion:

• Summarize contributions (NOT copy-paste from intro — rephrase)
• Limitations (be honest — reviewers appreciate this)
• Future work (1-2 concrete directions)
• Ethics statement and reproducibility statement (if venue requires)
• Target: 0.25-0.3 pages by default. Keep it as a concise close, not a second discussion section.

Appendix:

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

Step 3.5: Theory Paper Consistency Pass (theory papers only)

Run this pass after drafting all sections and before building the bibliography.

Trigger it when PAPER_PLAN.md labels the paper as theory/analysis, or when the drafted sections contain five or more formal result environments (theorem, lemma, proposition, or corollary).

Proof source search: search the workspace for standalone full-proof sources whose names or contents indicate a canonical proof version (proof, appendix, full, complete, supplement, supplementary). If one exists, ask:

Inline full proofs from {file}? [Y/n]

Default to Y. If accepted:

• import the full theorem/lemma statement plus proof block into the appendix source;
• use the main-body theorem statement as the canonical public statement;
• do not leave placeholders such as "see supplementary proof document" or "proof omitted for brevity";
• preserve theorem labels, equation labels, and proof structure exactly;
• keep main-body proof sketches short, but never let the appendix be sketch-only when a full proof source exists.

Restatement audit: compare every theorem/lemma/proposition statement restated in the appendix against the main-body version. Audit statements, hypotheses, case splits, quantifiers, domains, notation, variable names, and terminology for defined objects. Resolve all mismatches before Step 4.

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 only an arXiv clue is available, resolve it to DBLP, DOI/CrossRef, Semantic Scholar, OpenAlex, ACL Anthology, IEEE, ACM, or the official venue page before adding it to the final .bib
   • If not found and DBLP_BIBTEX = false, search only to identify the formal publication record; do not add arXiv BibTeX unless the user explicitly accepts an unpublished-preprint exception
   • 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 — published DOI preferred)

# If paper has a DOI, prefer the conference/journal DOI over an arXiv DOI.
curl -sLH "Accept: application/x-bibtex" "https://doi.org/{doi}"

Step C: Mark [VERIFY] (last resort) If DBLP, CrossRef, and other formal metadata sources return nothing, mark the entry with % [VERIFY] comment. Do NOT fabricate and do NOT silently fall back to arXiv in the final bibliography.

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.

If the DBLP/CrossRef flow is not enough, load ../../skills-codex/shared-references/citation-discipline.md for stricter fallback rules before adding placeholders.

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

import re
# 1. Grep all \citep{...}, \citet{...}, and \cite{...} from all .tex files
# 2. Extract unique keys (handle multi-cite like \citep{a,b,c} or \cite{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. Use published venue versions over arXiv preprints by default
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)
6. Treat arXiv entries as final-bibliography blockers unless explicitly accepted as unpublished-preprint exceptions.

Step 5: Evidence and Scientific Writing Quality Pass

Before language polish, run an evidence-story audit:

• Confirm the title, abstract, Results, and Discussion state the same bounded main finding; flag any stronger wording outside the storyboard.
• Check every primary figure/table reference, raw-result locator, version/date, sample size or definition of n, unit, group name, and statistic for consistency across prose, captions, tables, and source outputs.
• Confirm captions include the comparison and statistical details needed to interpret the claim, and that negative or contradictory results are visible or explicitly bounded rather than silently omitted.
• Check that methods identify sources/samples, key parameters, and an analysis method appropriate to the data type.
• Verify that citations support the claims they accompany, raw data remain traceable, and unresolved user decisions are listed rather than buried in prose.

Then run six sequential writing-quality passes. De-AI polish is included as one part of this quality pass, not a replacement for it.

Pass 1: Clutter Extraction — strip sentences to their cleanest components, remove filler, and remove AI-isms.

Pass 2: Active Voice and Verb Vitality — identify who did what, convert unnecessary passive voice, and resurrect smothered verbs.

Pass 3: Sentence Architecture — flag sentences over 40 words, keep subject and verb close, put familiar context first and new information later, and ensure each paragraph does one job.

Pass 4: Keyword Consistency — apply the Banana Rule: do not rename defined technical terms just to avoid repetition. If Methods defines a group, variable, or technique name, Results, Discussion, tables, and captions must use the same term.

Pass 5: Numerical and Citation Integrity — check sample sizes, percentages, significant figures, figure/table values, and whether citations support the claims they are attached to.

Pass 6: Anti-Defensive Writing — lead with the claim, result, or positive scope instead of what the paper does not claim, prove, cover, or attempt. Delete disclaimers that do not add evidence, scope, logic, conceptual precision, or necessary reader guidance. Keep necessary limitations once, calmly, in Methods, Discussion, Limitations, or the exact sentence whose interpretation depends on the limit. Convert negative scope into positive scope (The analysis focuses on X, not We do not claim to cover Y). Replace vague hedges (may, could, potentially) with precise evidence strength and scope; when uncertainty is real, name its source. Remove reflexive defense patterns unless the contrast is part of the argument: to be clear, it is worth noting, not X but Y, this should not be taken to mean, repeated however/although caveat openings.

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

First apply the sentence-level clarity rules from ../../skills-codex/shared-references/writing-principles.md:

• keep subject and verb close together
• put familiar context first and new information later
• place the most important information near the end of the sentence
• let each paragraph do one job
• use verbs for actions instead of nominalized nouns

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 GPT-5.4 xhigh:

spawn_agent:
  model: gpt-5.5
  reasoning_effort: xhigh
  message: |
    Review this [VENUE] paper draft (main body, excluding appendix).

    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 concise but sufficient, and do Related Work plus
       Conclusion stay within about 0.8 page for IEEE/Overleaf-style papers?
    6. For theory papers: are proof sketches adequate?
    7. Are figures/tables clearly described and properly referenced?
    8. Are there any solid-block paragraphs that nearly fill a two-column
       column and should be split?
    9. Does any high-impact prose sound defensive: repeated caveats, "we do
       not claim" framing, vague hedges, or limitation-first contribution
       paragraphs?

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

    [paste full draft text]

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 citation commands (\citep{}/\citet{} for ML conferences, \cite{} for IEEE) 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 + Conclusion are compact by default (about 0.8 page total for IEEE/Overleaf-style two-column papers unless explicitly overridden)
• No paragraph is a solid block that nearly fills a whole two-column column
• No unnecessary defensive writing remains in high-impact prose; necessary limitations are stated once in the right section
• references.bib contains ONLY cited entries (no bloat)
• No arXiv entries remain in the final .bib unless explicitly accepted as unpublished-preprint exceptions
• 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
• Venue-specific required sections/checklists satisfied (read ../../skills-codex/shared-references/venue-checklists.md if needed)
• A skim reader can recover the main claim from the title, abstract, introduction, and Figure 1/captions
• Title, abstract, Results, and Discussion agree on the same bounded main finding
• Every primary figure/table answers one storyboard question and is not used for an excluded conclusion
• Sample sizes/n, units, group names, captions, and statistics are consistent with raw-linked evidence
• Negative or contradictory results are visible or explicitly scoped; raw data and result versions are traceable
• User-facing interpretation or scope decisions remain listed in PAPER_PLAN.md

Key Rules

• 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)

• No over-claiming — use hedging language ("suggests", "indicates") for weak evidence

• Venue style matters — ML conferences (ICLR/NeurIPS/ICML) use natbib (\citep/\citet); IEEE venues use cite package (\cite{}, numeric). Never mix.

• Page limit rules differ by venue — ML conferences: main body to Conclusion, references/appendix NOT counted. IEEE: references ARE counted toward the page limit.

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

• Published-only bibliography by default — resolve arXiv clues to formal venue records and block final bibliography use of arXiv unless the user explicitly accepts an unpublished-preprint exception.

• 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

• Front-load the contribution — do not hide the payoff until the experiments or appendix

• Anti-defensive prose — state what the paper shows and where it applies; do not pad abstracts, introductions, contributions, captions, or conclusions with caveats aimed only at hypothetical objections.

Writing Quality Reference

• ../../skills-codex/shared-references/writing-principles.md — story framing, abstract/introduction patterns, sentence-level clarity, reviewer reading order
• ../../skills-codex/shared-references/venue-checklists.md — ICLR/NeurIPS/ICML/IEEE submission requirements to check before declaring done
• ../../skills-codex/shared-references/citation-discipline.md — stricter fallback for ambiguous citations

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.