Paper Writing Skill

Paper Writing Skill — Conference-Favorable Papers with Closed-Loop Verification

THE single skill for writing conference-favorable papers (写论文/投稿/改稿/camera-ready). It turns notes/outline/TeX project into a submission-quality LaTeX paper, applying the verified rhetoric playbook from arXiv 2608.08975 ("How Can Rhetoric Reward-Hack AI Reviewers?" — 4,200-manuscript controlled study of which rhetorical choices move AI reviewers), then runs a closed-loop verification script (scripts/verify_paper.py) until every check passes.

Inputs / Outputs

• Input: paper notes / outline / idea, or an existing TeX project directory (preferred when continuing a paper).
• Output: main.tex + appendix.tex (+ optional main_cn.tex Chinese twin via a Chinese-twin translation workflow), compiled PDFs, and verify_report.json from the closed-loop script.
• Work in a dedicated paper directory (e.g. /paper/), never in the repo root.

Phase 1 — Scaffold

1. Create main.tex (LaTeX article class, venue style file if required) and appendix.tex included via \input{appendix.tex}.
2. Required skeleton (the verify script enforces these):

• abstract environment (150–250 words)
• \section{Introduction}
• \section{Method} (or Approach | Model | Framework)
• \section{Experiments} (or Evaluation | Results | Empirical)
• \section{Conclusion} (or Conclusion and Discussion)
• Plus: Related Work, Limitations, Acknowledgments, \appendix.

1. Include a notation table in the method section and a figures/ dir for graphics.
2. Reference implementation to copy conventions from: a private bilingual paper project (main.tex/appendix.tex + a validate_publication.py-style layout check) — not included in this repo.

Phase 2 — Writing conventions

• Abstract 150–250 words; last sentence states the contribution.
• Introduction funnel: context → gap → contributions list.
• Every numeric claim is paired with a table/figure \ref or a \cite — no orphan numbers.
• Tables carry baseline rows and metric-direction arrows bound to the metric (CE$\downarrow$, BLEU$\uparrow$); never put arrows on data cells.
• Every training/result table is accompanied by its curve figure; captions state the arrow convention.
• Related Work ends with one differentiator sentence ("Unlike X, we ...").
• Limitations: honest, 3–4 items, author-identified future directions.
• No placeholder text ever (TODO/TBD/\ph{/XXX) — the verify script fails on them.
• Appendix evidence via the paper-appendix-evidence-tables workflow (macro-driven tables with pending-value fallbacks).

Phase 3 — Rhetoric playbook (verified from arXiv 2608.08975)

Mean ΔOA of rhetorical rewrites vs prompt-matched original (4,200 manuscripts, 120 ICLR 2026 submissions, 2 rewriters, 5 AI reviewers, standard+strict protocols):

| Dimension | Positive direction | Negative direction | |---|---|---| | Evidence framing | +.289 (largest contrast; weak-accept ±13pp; raises OA up to +0.93) | −.303 | | Novelty stance | +.187 | −.353 (penalty-driven — cautious framing is penalized; lowers OA up to −0.73) | | Scope framing | +.136 | −.289 (loss-driven under narrower framing) | | Contribution salience | +.192 | +.049 | | Technical register | +.142 | +.034 | | Linguistic complexity | +.017 | +.006 (don't over-invest) |

Rules derived from the study:

1. Evidence framing is the #1 lever: foreground reported comparisons, patterns, and comparative advantage in the results discussion. Make numbers visible. Understating comparisons costs ~0.3 OA.
2. Novelty stance is penalty-driven: appropriately cautious framing gets penalized. Assert supported claims and novelty confidently; avoid hedges ("we attempt", "preliminary evidence", "we believe").
3. Scope framing is loss-driven: frame scope broadly, but strictly within supported boundaries; don't tie claims narrowly to evaluated settings.
4. Signpost contributions (contribution salience) — explicit foregrounding helps slightly.
5. Starting-score dependence: rhetoric lifts low/mid drafts (joint-rewrite mean ΔOA +1.42 at initial AI score [1,3], +0.51 at [4,5]) and backfires on already-high ones (−0.88 at [8,10]). Weak-accept crossings cluster near OA 6 (accepted-ICLR mean 5.39). If the draft is weak/mid → invest in evidence+novelty framing; if already strong → minimize rhetorical noise, focus on substance.
6. Write for strict review: the strict protocol lowers mean OA by 1.36 (6.30 → 4.93) without changing rhetorical sensitivity. Every high rating must be backed by concrete evidence; presentation only counts when it changes the scientific assessment.
7. Revision-loop ceiling: gains concentrate by the 2nd pass; reviewer-guided rewriting does NOT beat an unguided second pass (guided − unguided: −0.008 … −0.067); joint rewriting is model-dependent (+.289 vs +.021). → Max 2 revision passes; never build elaborate multi-agent rewrite pipelines.
8. Caveat: LLM reviewers underperform at identifying substantive weaknesses — AI review supplements, never replaces, human sanity checks.

Operational definitions (from the official repo's rewrite prompts)

The paper's companion repo (github.com/MingLiiii/DissectingAIReviews, MIT) defines the six dimensions operationally. When polishing a paper, apply these semantics per dimension (positive direction; the negative is the opposite):

• Claim/novelty stance: rewrite existing claims, novelty statements, positioning, and contribution statements to read more assertive, confident, and decisive — but only where the original evidence supports the stronger wording; never turn limited evidence into unsupported certainty.
• Scope/generalization: broaden scope language (datasets, tasks, domains, model classes, assumptions, applications) within the boundaries already supported; keep the same actual conditions — change the framing, not the study.
• Quantitative evidence framing: present existing evidence, comparisons, metrics, and quantitative findings more clearly, prominently, and favorably — foreground comparison structure, margins, consistency, and practical significance; never create new results or change the meaning of reported values.
• Contribution structure: signpost contributions more explicitly, front-load them, make them scannable and reviewer-facing; surface supported contributions, never invent new ones.
• Technical register: make prose about technical objects, mechanisms, procedures, assumptions, and evaluation protocols more formal, precise, and specified; preserve notation meanings, formulas, algorithms, and results.
• Lexical complexity: use more sophisticated academic vocabulary and denser sentence structure — a prose-style intervention only; preserve what the paper reports.

Full-paper rewrite contract (applies to every dimension): rewrite ALL substantial prose paragraphs across every major section (abstract, introduction, method, experiments/results, discussion, conclusion) — captions, table notes, and result prose are in scope; you may reorganize sentences and revise framing; preserve the underlying claims, methods, datasets, settings, comparisons, metrics, and findings; never invent experiments, data, baselines, numbers, or unsupported conclusions; keep protected LaTeX structures valid (citation/label keys, URLs, graphics paths, displayed math environments, code/algorithm environments, bibliography files, custom macros, file set).

Phase 4 — Closed-loop verification (REQUIRED before every submission)

Script: /paper-writing-skill/scripts/verify_paper.py (stdlib-only Python 3.10+; run with python, or py -3 / a full interpreter path on Windows when python is not on PATH).

cd 
python "/paper-writing-skill/scripts/verify_paper.py"

The loop (this is the closed loop — do not skip it):

1. Run the script. Fix every FAIL item it reports (build errors, missing sections, placeholders, abstract length, unresolved refs/citations, citation count).
2. Re-run until exit code 0.
3. Optional AI-review simulation: --review (add --strict for the strict protocol, --gate-rating 6 to gate on rating ≥ 6 = weak-accept). Iterate on the reported weaknesses at most 2 passes (playbook rule 7), re-verifying after each pass.
4. Keep verify_report.json as the submission evidence.

CLI: verify_paper.py [--tex main.tex] [--no-build] [--review] [--strict] [--api-base URL] [--api-key KEY] [--model NAME] [--abstract-min 150] [--abstract-max 250] [--min-refs 12] [--gate-rating N] [--json verify_report.json]. API config via flags or env PAPER_REVIEW_API_BASE / PAPER_REVIEW_API_KEY / PAPER_REVIEW_MODEL (any OpenAI-compatible /chat/completions endpoint); without config the AI review is skipped gracefully and the deterministic checks still gate. Exit codes: 0 = no FAIL, 1 = ≥1 FAIL, 2 = usage error.

Checks performed: latexmk build + log parse (errors/undefined refs fail; overfull counts warn) · required sections · placeholders · abstract word count · label/citation resolution vs main.aux + min citation count · float references (warn) · numeric claims without backing refs (warn) · optional AI review in the paper's official ICLR-style form — rating ∈ {1,3,5,6,8,10}, soundness/presentation/contribution 1–4, confidence 1–5, ethics_flag, plus an additive rhetoric block (six 1–10 dimension ratings). The standard/strict prompts are the verbatim templates from github.com/MingLiiii/DissectingAIReviews (MIT).

Phase 5 — Optional

• Chinese twin: run a Chinese-twin translation workflow (xelatex+ctex, section-chunk translation, structural parity verification).
• Slides/video of the finished paper: hand off to the paper-to-beamer / paper-slides-to-video pipelines — do NOT do it inside this skill.

Companion resources

• Chinese-twin translation workflow — CN twin build (xelatex+ctex).
• paper-appendix-evidence-tables workflow — macro-driven appendix evidence tables.
• A private bilingual paper project — reference implementation for EN/CN twin, layout verification, metric-arrow conventions (not included).
• A top-conference writing-guide course deck — P1–P15 innovation-pattern knowledge (what makes papers Oral-caliber; not included).
• github.com/MingLiiii/DissectingAIReviews — the paper's official companion repo (MIT): the exact standard/strict review prompts and the six-dimension rewrite prompt bank that this skill's prompts and playbook are derived from.

Windows gotchas

• Patch .tex via write-based scripts (bash heredocs mangle backslashes).
• Detect failed compiles by grepping the log for "Output written", not by file mtime — a stale main.pdf can mask a failed run.
• python may not be on PATH on Windows — use py -3 or the full interpreter path.
• \（/\） escaped fullwidth punctuation are undefined control sequences in xelatex — never write them.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: yhbcode000
• Source: yhbcode000/paper-writing-skill
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.