本文へ移動
cccskills
無料GitHub で公開

paper-self-revise

Revise an academic paper based on internal review comments. Reads review report or annotated manuscript, then applies revisions one by one with user approval. Trigger when user says "self revise" / "paper-selfrevise" / "内部修改" / "根据审稿意见修改".

インストール方法を見る

含まれるファイル(1)

  • SKILL.md14.3 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Paper Self-Revise

Overview

This skill revises an academic paper (main.tex) based on internal review comments. The user provides a project folder containing the manuscript and review materials. The skill extracts review comments, then processes them one by one: displaying the comment, showing the original text, proposing a revision, and asking for user approval before applying the change.

Input: A project folder containing:

  • main.tex — the main manuscript (the file to be modified)
  • Subfolders with figures/tables (referenced by the manuscript)
  • Review materials (one or more of the following):
    • A review report file (e.g., review.txt, review.md, review.pdf, report.txt, report.md, report.pdf)
    • An annotated manuscript (e.g., annotated.tex, annotated.pdf) where annotations/comments serve as review comments

Output: The revised main.tex with all approved changes applied in place.


Workflow

Phase 0: Review Report Check

  1. Receive folder path from user (via $ARGUMENTS or ask).
  2. Ask the user: "您是否已有现成的审稿报告(review report)?(yes/no)"
  3. If yes: Proceed to Phase 1.
  4. If no: Draft a review report before proceeding: a. Read main.tex and all supporting materials (tables, figures) in the project folder to fully understand the paper. b. Draft a review report in English (~3000 words), structured as:
    • A brief overall assessment (1–2 paragraphs)
    • Numbered comments: Comment 1, Comment 2, Comment 3, etc.
    • Each comment should identify a specific issue, explain why it matters, and suggest how to address it.
    • IMPORTANT: Exclude any comments related to empirical tests that would require changing software code (e.g., re-running regressions, adding new specifications, changing estimation methods, running robustness checks in Stata/R/Python). Focus only on writing, argumentation, framing, literature, structure, clarity, interpretation, and presentation. c. Save the review report as review.pdf in the project folder using a Bash command to generate the PDF (e.g., via pandoc or a Python script with reportlab/fpdf). If pandoc is available, first write the content to review.md then convert: pandoc review.md -o review.pdf. If pandoc is not available, use Python to generate the PDF. d. Count the words of the generated review report (count words in the source text — the review.md content or the raw English text used to build review.pdf, excluding markdown/PDF formatting markup). Use a Bash command such as wc -w review.md or a Python word-count snippet on the text. e. Inform the user, displaying the actual word count: "审稿报告已生成并保存为 review.pdf(共 X 词),现在开始逐条修改。" — substitute X with the measured count. f. Proceed to Phase 1 using the newly created review.pdf as the review material.

Phase 1: Initialization

  1. Scan the folder using Glob to identify all files:
    • Locate main.tex (the target file to revise).
    • Locate review materials by searching for files matching common review file patterns: review*, report*, referee*, comments*, annotated*, feedback*, revision*.
    • If multiple review files are found, list them and ask the user which one(s) to use.
    • If no review file is found, ask the user to specify the review file path.
  2. Read main.tex in full to understand the current manuscript (skip if already read in Phase 0).
  3. Read the review file(s):
    • For .tex files: look for comments marked with % or annotation commands (e.g., \annotate{}, \comment{}, \todo{}, or any custom annotation commands).
    • For .txt / .md files: read the full content as review comments.
    • For .pdf files: read the PDF content to extract review text.
    • For .docx files: use Bash to extract text content.
  4. Detect tracked changes (revision marks) in the annotated manuscript:
    • Look for revision markup such as strikethrough/deletion marks, inserted text, colored text indicating additions/deletions, or LaTeX track-changes commands (e.g., \added{}, \deleted{}, \replaced{}{}, \st{}, \sout{}, \hl{}, or similar commands from packages like changes, trackchanges, latexdiff).
    • Tracked changes in the body text are revision items just like annotation comments — each one must be presented to the user for approval in Phase 3.

Phase 2: Extract and Structure Review Comments

  1. Parse the review materials to extract individual review comments/suggestions. This includes two categories:
    • Annotation comments: margin comments, %-comments, \todo{}, \comment{}, etc.
    • Tracked changes: inline revision marks (insertions, deletions, replacements) in the body text. Each distinct tracked change is treated as a separate revision item, equivalent to an annotation comment.
  2. Filter out formatting-only comments: Automatically discard any comments or tracked changes that are purely about formatting (e.g., font size, spacing, margins, line breaks, paragraph indentation, column layout, page layout, figure/table placement commands). Do NOT present these to the user. Only retain comments about substantive content (wording, argumentation, data, analysis, structure, logic, references, etc.).
  3. Structure each comment as a discrete revision item with:
    • Comment number (sequential)
    • Type — either "批注" (annotation comment) or "修订痕迹" (tracked change)
    • Review comment text (the reviewer's original words, or for tracked changes: a description of what was deleted/inserted/replaced)
    • Location in manuscript (section, paragraph, or specific text the comment refers to)
  4. Present the full list of comments to the user:
I have extracted the following review comments:

No. | Type     | Location                  | Comment Summary
----|----------|---------------------------|------------------------------------------
1   | 批注     | Section 3, para 2         | Clarify the identification strategy
2   | 修订痕迹 | Section 4.1, Table 2      | Replace "effect" with "impact" in sentence about...
3   | 批注     | Section 5, para 1         | Strengthen the mechanism discussion
...

Total: N comments. Shall I proceed with revisions one by one?
  1. Wait for user confirmation before proceeding.

Phase 3: Iterative Revision (Core Loop)

For each review comment, follow this exact interaction pattern:

Step 1: Display the review comment

For annotation comments (批注):

───────────────────────────────────────────
批注 [X/N]:
[Full text of the reviewer's comment]
───────────────────────────────────────────

For tracked changes (修订痕迹):

───────────────────────────────────────────
修订痕迹 [X/N]:
[Description: e.g., "删除了 'XXX',替换为 'YYY'" or "插入了 'XXX'" or "删除了 'XXX'"]
───────────────────────────────────────────

Step 2: Show original text

Locate and display the relevant passage from main.tex:

Original text:
┌─────────────────────────────────────────┐
│ [The exact paragraph(s) from main.tex   │
│  that this comment refers to]           │
└─────────────────────────────────────────┘

If the comment refers to missing content (e.g., "add a discussion of X"), show the surrounding context where the new text should be inserted.

Step 3: Propose revision

For annotation comments: Generate a revised version of the text that addresses the reviewer's comment.

For tracked changes: The proposed revision is the result of accepting the tracked change (i.e., applying the deletion/insertion/replacement as marked by the reviewer).

Proposed revision:
┌─────────────────────────────────────────┐
│ [The revised paragraph(s) that address  │
│  the reviewer's comment]                │
└─────────────────────────────────────────┘

Rationale: [Brief explanation of what was changed and why]

Step 4: Ask for user decision

Use AskUserQuestion to ask:

Do you approve this revision? (yes / no / skip / edit)
- yes: Apply this revision to main.tex
- no: I will propose a new revision
- skip: Skip this comment and move to the next
- edit: You provide the revised text yourself

Step 5: Handle user response

  • "yes" / "y" / "同意" / "好": Use the Edit tool to replace the original text with the proposed revision in main.tex. Confirm the change was applied, then move to the next comment.
  • "no" / "n" / "不同意" / "不行": Ask the user what they would like changed about the proposed revision, then generate a new proposal. Return to Step 3.
  • "skip" / "跳过": Skip this comment without making changes. Move to the next comment.
  • "edit" / "我来改": Ask the user to provide their own revised text via AskUserQuestion, then apply that text using the Edit tool. Move to the next comment.

Phase 4: Completion

After processing all comments:

  1. Present a summary:
Revision complete!

Applied: X / N items (A annotation comments + B tracked changes)
Skipped: Y items
Auto-filtered (formatting-only): Z items
Total changes made to main.tex: X

Modified sections:
- Section 3: 2 revisions
- Section 4: 1 revision
- Section 5: 3 revisions
  1. Offer to re-read the revised main.tex to check for consistency or flow issues introduced by the revisions.

Writing Rules

Style and Formatting

  • Maintain the same writing style as the original manuscript (do not change the overall tone or voice).
  • Preserve all LaTeX formatting: \section{}, \subsection{}, \ref{}, \cite{}, \citep{}, math environments, etc.
  • Preserve all labels (\label{}), cross-references (\ref{}), and citations intact unless the review specifically asks to change them.
  • When adding new citations, use the same citation style as the rest of the paper (typically 30% \cite{} / 70% \citep{}).
  • When reporting regression results, use t-statistics in parentheses and "significant at X% level" — no p-values or confidence intervals.
  • Avoid the em-dash (---): Do not use the LaTeX em-dash --- in revised or newly written prose. It is a common AI-writing tell and clutters academic text. Replace each occurrence with the punctuation its grammatical role calls for: a parenthetical aside becomes parentheses (or commas if short); a connector before an explanation becomes a colon or semicolon; an abrupt shift becomes a period and two sentences; a list continuation becomes a comma plus a connecting word. Leave the en-dash -- used for page and number ranges untouched (e.g., pages={259--271}, 2012--2026). When a revision touches a paragraph, also strip any pre-existing --- from that paragraph as part of the edit.

Revision Principles

  • Minimal changes: Only modify what is necessary to address the review comment. Do not rewrite entire paragraphs when a sentence-level edit suffices.
  • Preserve structure: Do not reorganize sections or move paragraphs unless the review specifically requests it.
  • No fabrication: Never invent data, results, or citations. If the review asks for additional analysis not present in the tables/figures, flag this to the user rather than making up results.
  • Consistency: After each revision, ensure the changed text is consistent with surrounding paragraphs (transitions, terminology, notation).
  • Scope discipline: Only address what the reviewer asked for. Do not make additional "improvements" beyond the scope of each comment.

Handling Special Cases

  • Comment refers to a table/figure: Read the relevant table/figure .tex file from the project folder before proposing the revision, to ensure accuracy.
  • Comment asks for new content: Propose where to insert it (after which paragraph/section) and draft the new text.
  • Comment asks to delete content: Show what will be removed and confirm before deleting.
  • Comment is vague or unclear: Present your interpretation to the user and ask for clarification before proposing a revision.
  • Comment requires changes in multiple locations: Handle each location as a sub-item (e.g., Comment 3a, 3b, 3c) within the same comment block.
  • Comment about bibliography/references: If changes involve ref.bib or similar bibliography files, modify those files as well.
  • Tracked changes (revision marks): Inline deletions, insertions, or replacements marked in the annotated manuscript are treated identically to annotation comments. Present each tracked change to the user with the original text and the proposed result of accepting the change. The user decides whether to accept or reject each one.
  • Formatting-only comments: Comments or tracked changes that are purely about formatting (font, spacing, margins, line breaks, indentation, column layout, page layout, float placement, \vspace, \hspace, \newpage, \clearpage, figure/table positioning options like [htbp], etc.) must be automatically skipped without presenting to the user. Only substantive content comments are shown.

Language

  • Communicate with the user in Chinese (the user's preferred language for interaction).
  • Write all LaTeX content and revisions in English (academic paper language).
  • Review comment display should preserve the original language of the review.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

a1

無料

VS-Enhanced Research Question Refiner - Prevents Mode Collapse and derives differentiated research questions Enhanced VS 3-Phase process: Modal question avoidance, alternatives presentation, differentiated RQ recommendation Use when: refining research ideas, formulating research questions, clarifying scope Triggers: research question, 연구 질문, PICO, SPIDER, research idea

日本語の概要は準備中です。原文の説明を表示しています。

brycewang-stanford/Auto-Empirical-Research-Skills4,6032026年10月12日 更新

a2

無料

VS-Enhanced Theoretical Framework Architect with Critique & Visualization Full VS 5-Phase process: Modal theory avoidance, Long-tail exploration, differentiated framework presentation Absorbed A3 (Devil's Advocate) critique and A6 (Conceptual Framework Visualizer) capabilities Use when: building theoretical foundations, designing conceptual models, deriving hypotheses, critiquing frameworks, visualizing models Triggers: theoretical framework, 이론적 프레임워크, conceptual model, 개념적 모형, hypothesis derivation, critique, devil's advocate, 반론, visualization, diagram

日本語の概要は準備中です。原文の説明を表示しています。

brycewang-stanford/Auto-Empirical-Research-Skills4,6032026年10月12日 更新

a5

無料

Agent A5 - Paradigm & Worldview Advisor - Philosophical foundations for research design. Covers ontology, epistemology, axiology, and methodology alignment. Use when: establishing philosophical foundations, justifying methodological choices, writing positionality statements Triggers: paradigm, 패러다임, ontology, epistemology, worldview, 세계관, philosophical foundations, 철학적 기초

日本語の概要は準備中です。原文の説明を表示しています。

brycewang-stanford/Auto-Empirical-Research-Skills4,6032026年10月12日 更新

Academic paper writing skill with 12-agent pipeline. v2.5: Style Calibration (learn author's writing voice from past papers) + Writing Quality Check (writing quality checklist for natural prose). Supports IMRaD, literature review, theoretical, case study, policy brief, and conference paper structures. APA 7.0 (default), Chicago, MLA, IEEE, Vancouver citation formats. Bilingual abstracts (zh-TW + EN). Multi-format output (LaTeX, DOCX, PDF, Markdown). Triggers on: write paper, academic paper, paper outline, write abstract, revise paper, check citations, convert to LaTeX, guide my paper, parse reviews, revision roadmap, 寫論文, 學術論文, 論文大綱, 寫摘要, 修改論文, 檢查引用, 引導我寫論文, 帶我規劃論文, 逐章規劃, 論文架構, 審查意見, 修訂路線圖.

日本語の概要は準備中です。原文の説明を表示しています。

brycewang-stanford/Auto-Empirical-Research-Skills4,6032026年10月12日 更新

Systematic writing framework for philosophy and interdisciplinary academic papers from optimized outline to submission-ready manuscript. Use when users want to: (1) write a paper from a detailed outline, (2) ensure quality control during writing, (3) maintain consistency across chapters, (4) prepare a submission-ready manuscript, or (5) systematically execute a planned paper. Triggered by phrases like 'write the paper from this outline,' 'compose the full manuscript,' 'execute the outline,' or when users have completed strategic planning (academic-paper-strategist skill) and are ready to write. Takes optimized outline as input; outputs complete manuscript with iterative quality checks.

日本語の概要は準備中です。原文の説明を表示しています。

brycewang-stanford/Auto-Empirical-Research-Skills4,6032026年10月12日 更新

Multi-perspective academic paper review with dynamic reviewer personas. Simulates 5 independent reviewers (EIC + 3 peer reviewers + Devil's Advocate) with field-specific expertise. Supports full review, re-review (verification), quick assessment, methodology focus, and Socratic guided modes. Triggers on: review paper, peer review, manuscript review, referee report, review my paper, critique paper, simulate review, editorial review.

日本語の概要は準備中です。原文の説明を表示しています。

brycewang-stanford/Auto-Empirical-Research-Skills4,6032026年10月12日 更新

brycewang-stanford のスキルをすべて見る

このスキルの問題を報告する