Skip to content
Back to skills

Paper Production

ASecurity

End-to-end paper production pipeline fusing the best writing skills — idea gate (idea-evaluator), skeleton (tech/benchmark-paper-template), Introduction logic chain (intro-drafter), evidence arc (paper-narrative), figure design (figure-style/figure-composer/academic-plotting), drafting with citations (paper-writing + ml-paper-writing venue templates), dual review (reviewer agent + pre-submission-reviewer). Use when the user wants to write a paper from research results, structure a manuscript ...

  • 9 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 7, 2026
research

Security analysis

A100/100

Scanned October 7, 2026

npx -y skills add Yichen-ZJU/claude-for-research --skill paper-production --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Paper Production?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Paper Production
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/yichen-zju-paper-production/badge)](https://www.skillsdirectory.com/skills/yichen-zju-paper-production)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: paper-production
description: End-to-end paper production pipeline fusing the best writing skills — idea gate (idea-evaluator), skeleton (tech/benchmark-paper-template), Introduction logic chain (intro-drafter), evidence arc (paper-narrative), figure design (figure-style/figure-composer/academic-plotting), drafting with citations (paper-writing + ml-paper-writing venue templates), dual review (reviewer agent + pre-submission-reviewer). Use when the user wants to write a paper from research results, structure a manuscript for a specific venue, or run a full idea-to-submission pipeline.
argument-hint: <topic-or-results-dir> [--venue <venue>] [--type technique|benchmark|new-problem]
---

# Paper Production(论文生产总装线)

融合各来源最强写作 skills 的端到端论文管线。每个阶段一个最强工具,阶段间有质量门——不过门不进入下一阶段。

产物落盘遵循全局约定:`papers/<slug>.md`(或 LaTeX 项目目录)、计划入 `outputs/.plans/`、关键阶段留 provenance。

## 管线

### Stage 0 — 选题闸门(可跳过但强烈建议)

调 `idea-evaluator`:五维(Higher/Faster/Stronger/Cheaper/Broader)+ 想法生命周期 + 能力匹配 + 致命缺陷审计,输出 reviewer 式判决。

**质量门**:判决为"不值得做"或存在致命缺陷 → 停下来把判决交给用户,不硬写。

### Stage 1 — 逻辑骨架

- 技术/新方法论文 → `tech-paper-template`(思考模板表:背景→局限→关键思想→挑战→方法模块→贡献,Technique vs New Problem/Setting 定位,四点自洽检查)
- Benchmark/评测论文 → `benchmark-paper-template`(五支柱:Research Gap / Construction Pipeline / Evaluation Framework / Empirical Findings / Companion Method)

**质量门**:四点自洽检查通过(挑战↔方法模块↔贡献一一对应)。

### Stage 2 — Introduction

调 `intro-drafter`:从结构化 Flowchart 产六段大纲(背景+running example → 现有局限 → 问题本质与目标 → 关键挑战 → 方案总览 → 贡献列表),贡献与挑战对齐。

**质量门**:六段逻辑链能一口气讲通;每段都能在骨架里找到对应。

### Stage 3 — 全文叙事与图规划

调 `paper-narrative`:提炼中心 claim、证据链、图的顺序(每张图支撑哪个决策),标出"图 claim 超出数据证据"的危险点。
图的设计与绘制走 `figure-style` / `figure-composer` / `academic-plotting`。

**质量门**:每张图有明确的"它证明什么";没有无数据支撑的图。

### Stage 4 — 初稿

走 `paper-writing` 工作流:`writer` agent 从研究文件起草(不编造结果,缺证据留 TODO)→ `verifier` agent 逐条加引用、核验 URL、删无源声明。
排版模板:ML 会议(NeurIPS/ICML/ICLR/ACL/AAAI/COLM)用 `ml-paper-writing` 的 LaTeX 模板;系统会议(OSDI/SOSP/ASPLOS/NSDI)用 `systems-paper-writing`。
**文笔**:段落和句子层面遵循 `research-paper-writing`(彭思达方法论:一段一信息、首句点题、句间流转),各节 prose 指导读它的 `references/{abstract,introduction,related-work,method,experiments,conclusion}.md`。若 `outputs/style-profile-<author>.md` 存在(由 `style-calibration` 建立),writer 起草时把它作为软约束贴合作者声纹。

**质量门**:verifier 的 result-provenance 审计无残留无源数字/图/表;每节写完做 reverse outlining 检验(topic sentence → thesis 映射干净)。

### Stage 5 — 双审查(两个都要过)

1. **证据严谨** → `reviewer` agent:FATAL/MAJOR/MINOR,unsupported claims、zombie sections、单源关键声明。FATAL 必须修,修完再审一轮。
2. **写作品味** → `pre-submission-reviewer`:宏观逻辑/写作细节/语法/LaTeX/图质量 + AI 腔词汇 + 破折号滥用。段落流畅度存疑时,用 `research-paper-writing` 的 Paragraph Clarity Check(外部读者视角 + reverse outlining)复核。

**质量门**:无 FATAL、无 CRITICAL;MAJOR 已修或显式记录为 known limitation。

### Stage 5.5 — AI 披露(必做)

调 `ai-use-disclosure`:按目标 venue 政策生成 AI 使用声明(并记录政策检查日期)。披露范围 = 实际使用范围。缺这一步不交付。

### Stage 6 —(可选)演讲

`presenting-conference-talks`:Beamer  slides + 讲稿。

## 用法

- **全流程**:"把 experiments/ 里的结果写成 NeurIPS 论文" → 从 Stage 1 开始(Stage 0 可选)
- **中途进入**:用户已有骨架/初稿 → 从对应阶段切入,之前的阶段视为已过门但需快速 sanity check
- **单独阶段**:用户只要 intro / 只要审稿 → 直接调对应 skill,不必走全线

## 纪律

- 每个 Stage 完成时在 `outputs/.plans/<slug>-paper.md` 里记录:阶段、所用 skill、质量门结果、遗留问题。
- 诚实标注:blocked/unverified/inferred 不许粉饰(全局 provenance 约定)。
- AI 腔和模板腔是交付缺陷,不是风格问题 —— Stage 5.2 不过不许交付。

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.