Skip to content

Haojae/scipilot-writing-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

scipilot-writing-skill

SciPilot Skills family. Academic writing & polishing co-pilot — a top-journal editor and a harsh reviewer, built into your workflow. SciPilot Skills 家族成员 — 学术论文写作与润色副驾驶:把"顶刊编辑 + 严苛审稿人"装进你的写作流程。

License: MIT Python: 3.9+ Status: v1.0.0 Mode Lint Claude Code Skill

A Claude Code / Codex / Cursor Skill that takes scientific text the last mile: from a Chinese draft, a rough paragraph, or a whole section, to submission-grade English (or Chinese) prose. It covers translation, condensing, expanding, polishing, de-AI / humanizing, logic checking, section drafting (Title / Abstract / Introduction-CARS / Methods / Results / Discussion), figure & table captions, results analysis, reviewer-style self-check, cover letters, and rebuttals. Unlike a prompt cookbook, writing quality here is a machine-checked contract: every delivery passes writing_lint.py (deterministic AI-tell / LaTeX-escape / Markdown-contamination / full-width-punctuation checks) plus an AI read-back self-review loop.

中文文档 | English


中文文档

概览

科研写作最磨人的,往往不是"不会写",而是反复调润色 prompt、改完不知道好没好、投出去才发现满是 中文乱标点 / LaTeX 没转义 / 读着一股 AI 味。scipilot-writing-skill 是 SciPilot 家族的写作成员, 专治这"最后一公里"——先判断后下笔,宁缺毋滥,改完必自检。

为什么这不只是一份 prompt 清单

普通写作 prompt 清单          scipilot-writing-skill
──────────────────         ──────────────────────
复制 prompt → 粘贴 →  改完   先按 Stage 0 问清:任务/载体/目标刊/语言/学科/保守度
凭感觉觉得"好像不错"          改写走三段式:改后文 + 回译核对 + 修改日志
                            机器自检 writing_lint:AI 指纹/转义/全角/被动/节奏
                            AI 读稿审稿人视角自审:逻辑/夸大/AI味/一致性 → 回改
                            诚信红线:只动表达不动数值/方向/结论,绝不臆造引用

写作质量证据链

这是 SciPilot 家族的统一签名——把"写得好"从口头承诺变成机器可核对的记录,与 scipilot-figure-skill 的视觉自检闭环、scipilot-cite-skill 的幻觉门控同源。

改写 → writing_lint.py 机器自检(确定性指纹)
        + AI 读稿审稿人自审(感知/逻辑问题)
                  ↓ 发现问题
        回改 → 重检,最多 3 轮,残留如实标注
  • 机器抓确定性问题writing_lint.py 命中即报——中英 AI 指纹词、机械连接词、悬垂 -ing、 否定式平行、破折号密度、空泛归因、模型名所有格、LaTeX 未转义、Word 端残留 Markdown、 中文半角标点、被动比例、句长节奏(burstiness)。输出 lint_report.json
  • AI 读稿抓感知问题:节奏是否机械、是否 overclaim、术语是否漂移——这些程序查不出。
  • 诚信复核:diff 确认没动任何数值、公式、引用键、结论方向。

核心工作流(Stage 0–7)

0. 信息收集 ── 任务/载体/目标刊/语言/学科/保守度(必问,禁止默默假设)
   ↓
1. 解析载体 ── .tex / .docx / .md / 纯文本,定位处理范围
   ↓
2. 选路由   ── prompt_library + section_playbooks + journal_styles
   ↓
3. 起草改写 ── 三段式:改后文 + 回译核对 + 修改日志
   ↓
4. 机器自检 ── writing_lint.py(Gate,命中 FAIL 必处理)
   ↓
5. AI 自审  ── 审稿人视角读稿,发现问题回 3,最多 3 轮
   ↓
6. 一致性诚信复核 ── 术语/缩写/时态/数字/US-UK;diff 确认未动事实
   ↓
7. 交付报告 ── 做了什么 + lint 摘要 + 残留项 + 家族协作建议

七条铁律(IRON RULES)

  1. 诚信:只动表达,绝不改数据/数值/方向/结论,绝不臆造文献或结果。
  2. 宁缺毋滥:原文已好就保留并肯定,不为换词而换词。
  3. 透明可溯:绝不静默重写,永远给三段式(改后文 + 回译 + 修改日志)。
  4. 去 AI 味但留人味:删 AI 指纹,但保留作者观点、节奏与术语。
  5. 格式纯净随载体:LaTeX 转义、Word 全角纯文本、拒绝无谓列表化。
  6. 断言=证据:hedging/boosting 匹配证据强度,绝不 overclaim,绝不把相关说因果。
  7. 机检门控:交付前必跑 writing_lint.py,FAIL 必处理,不伪造检查。

覆盖的写作任务(同类清单的全部 + 更多)

中译英润色 · 英译中 · 中文重写 · 缩写 · 扩写 · 英文润色 · 中文润色 · 逻辑检查 · 去 AI 味(中/英) · 图标题 · 表标题 · 实验结果分析 · 审稿人视角自检 —— 以上对齐同类 prompt 清单; 额外:Title 创作 · Abstract 起草 · Introduction(CARS) · Methods/Results/Discussion 起草 · Cover letter · Rebuttal / 审稿回复 —— 高水平 SCI 高频、同类清单普遍缺失的能力。

安装

请帮我安装这个 Skill:https://github.com/Haojae/scipilot-writing-skill.git

或手动:

git clone https://github.com/Haojae/scipilot-writing-skill.git \
          ~/.claude/skills/scipilot-writing-skill
# 仅处理 .docx 时才需要:
pip install python-docx

writing_lint.py / latex_clean.py / text_stats.py 只用 Python 标准库,开箱即用。

命令行直接调脚本

# 写作质量机检(核心)
python scripts/writing_lint.py draft.tex --mode latex --lang en --report lint_report.json
python scripts/writing_lint.py draft.txt --mode word  --lang zh

# LaTeX 转义 + 体检
python scripts/latex_clean.py draft.tex --check
python scripts/latex_clean.py draft.tex            # 转义后输出

# 量化指标:句长节奏 / 被动 / 名词化 / hedging
python scripts/text_stats.py draft.txt --lang en

# 读 Word 为纯文本(喂给机检)
python scripts/docx_text.py read paper.docx --to body.txt

SciPilot Skills 家族

Skill 状态 功能
scipilot-cite-skill v1.0.0 文献检索与引用插入
scipilot-figure-skill v2.1.0 可视化顾问 + 绘制 + 视觉自检闭环
scipilot-writing-skill v1.0.0 (本仓库) 写作与润色 + 写作质量证据链
scipilot-review-skill 规划中 AI 模拟审稿
scipilot-submit-skill 规划中 投稿格式适配
scipilot-read-skill 规划中 论文阅读与翻译

许可证

MIT © 2026 Haojae


English

Overview

The hardest part of scientific writing is rarely "I can't write" — it's re-tuning polish prompts forever, never knowing if the result is good, and discovering at submission time that the text is full of half-width punctuation, unescaped LaTeX, or an unmistakable AI smell. scipilot-writing-skill is the writing member of the SciPilot family, built for that last mile: judge before you write, change only what needs changing, and always self-check.

Why this isn't just a prompt cookbook

Generic prompt list          scipilot-writing-skill
──────────────────          ──────────────────────
copy prompt → paste → edit   First asks: task / medium / target journal / language / field / how aggressive
"looks fine I guess"         Rewrites in 3 parts: revised text + back-translation + change log
                             Machine self-check (writing_lint): AI-tells / escapes / passive / rhythm
                             AI read-back as a reviewer: logic / overclaim / AI smell / consistency
                             Integrity line: never touch numbers / direction / conclusions; never fabricate refs

Writing-quality evidence chain

The family signature — turning "it reads well" from a verbal promise into a machine-checkable record, the same DNA as the visual self-check loop in scipilot-figure-skill and the hallucination gate in scipilot-cite-skill.

  • Machine catches deterministic issues: writing_lint.py flags AI-tell words (EN+ZH), mechanical connectives, dangling -ing, negative parallelism, em-dash density, vague attribution, possessive on model names, unescaped LaTeX specials, Markdown contamination in Word output, half-width punctuation in Chinese, passive ratio, and sentence-length burstiness. Emits lint_report.json.
  • AI catches perceptual issues: monotone rhythm, overclaiming, terminology drift — invisible to a linter.
  • Integrity review: a diff confirms no number, formula, citation key, or conclusion was altered.

Tasks covered (everything a prompt list does, and more)

zh→en polish · en→zh · Chinese rewrite · condense · expand · English polish · Chinese polish · logic check · de-AI (zh/en) · figure caption · table caption · results analysis · reviewer self-check — all on par with comparable prompt collections. Plus: Title ideation · Abstract drafting · Introduction (CARS) · Methods/Results/Discussion drafting · Cover letter · Rebuttal — high-value for SCI submission and usually missing from prompt lists.

Installation

Please install this Skill for me: https://github.com/Haojae/scipilot-writing-skill.git
git clone https://github.com/Haojae/scipilot-writing-skill.git \
          ~/.claude/skills/scipilot-writing-skill
pip install python-docx   # only needed for .docx I/O

SciPilot Skills family

Skill Status Purpose
scipilot-cite-skill v1.0.0 Reference discovery & insertion
scipilot-figure-skill v2.1.0 Visualization advisor + renderer + visual self-check
scipilot-writing-skill v1.0.0 (this repo) Writing & polishing + writing-quality evidence chain
scipilot-review-skill Planned AI peer-review simulation
scipilot-submit-skill Planned Submission formatting
scipilot-read-skill Planned Paper reading & translation

License

MIT © 2026 Haojae

依赖 / Dependencies

# 核心脚本零依赖(Python 3.9+ 标准库)
python-docx>=1.1   # optional; only for reading/writing .docx

About

SciPilot family: academic writing and polishing co-pilot, with a machine-checked writing-quality evidence chain (AI-tell linter + reviewer self-check loop).

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages