| name | manuscript-review-revise |
| description | AI-powered manuscript review and revision system inspired by APRES (ICLR 2026). Evaluates scientific manuscripts using ScholarEval 8-dimension rubric plus citation-predictive heuristics, then performs targeted revisions while preserving core scientific claims. Outputs before/after comparison with improvement metrics. Use when the user says "/review", "帮我审一下", "review my manuscript", "improve this paper", "polish this draft", or provides a manuscript for quality improvement. Also triggered by "审稿", "修改论文", "润色". |
Manuscript Review & Revision System
Evaluate and improve scientific manuscripts through a closed-loop process: Score → Identify Weaknesses → Revise → Re-score → Compare. Inspired by APRES (Meta Superintelligence Labs, ICLR 2026).
Core Principles
- Never modify core scientific claims — preserve all data, results, and conclusions
- Never add unverified data or fabricated citations — only restructure existing content
- Every revision has a stated reason — traceable, auditable changes
- Quantitative before/after comparison — ScholarEval scores pre and post revision
- Improve presentation, not science — clarity, structure, flow, completeness
When to Use
- User says
/review or /review <path-to-manuscript>
- User asks "帮我审一下这篇论文" or "review my manuscript"
- User asks "润色" or "polish this draft"
- After ScienceClaw generates a research report, offer: "需要用审修系统优化这份报告吗?"
Workflow
Phase 1: ScholarEval Assessment (8 Dimensions)
Score the manuscript on each dimension (0.00–1.00):
| Dimension | Weight | Evaluation Criteria |
|---|
| Novelty | 15% | Does this advance knowledge? Are claims clearly differentiated from prior work? |
| Rigor | 25% | Methodology sound? Statistics correct? Controls adequate? Sample sizes reported? |
| Clarity | 10% | Writing clear? Figures self-explanatory? Logical flow between sections? |
| Reproducibility | 15% | Methods detailed enough to replicate? Software versions stated? Data accessible? |
| Impact | 20% | Does this matter for the field? Broad or narrow implications? |
| Coherence | 10% | Do all parts fit together? Introduction → Methods → Results → Discussion aligned? |
| Limitations | 3% | Are limitations honestly acknowledged? Not buried or trivialized? |
| Ethics | 2% | Ethical standards met? IRB mentioned if applicable? Conflicts disclosed? |
Compute weighted average. Output initial verdict: accept (≥0.75), minor_revision (≥0.60), major_revision (≥0.40), reject (<0.40).
Phase 2: Citation-Impact Heuristics
Evaluate 10 presentation factors that predict higher citation impact:
| # | Factor | Check | Weight |
|---|
| 1 | Title specificity | Contains key finding or quantitative result? Not vague? | High |
| 2 | Abstract conclusion | Has a clear, quantitative take-home message? | High |
| 3 | Figure self-sufficiency | Can each figure be understood from its caption alone? | High |
| 4 | Methods reproducibility | Software versions, parameters, thresholds all stated? | Medium |
| 5 | Statistical reporting | Effect sizes + CIs alongside p-values? Test assumptions verified? | Medium |
| 6 | Discussion balance | Presents counter-arguments and alternative interpretations? | Medium |
| 7 | Limitations honesty | Dedicated section with specific (not generic) limitations? | Medium |
| 8 | Introduction funnel | Narrows from broad context → gap → specific question? | Low |
| 9 | Reference recency | Includes papers from last 2 years? Not relying on outdated reviews? | Low |
| 10 | Data availability | States where data/code can be accessed? | Low |
Score each 0–1. Flag factors scoring below 0.5 as revision targets.
Phase 3: Generate Revision Plan
Sort all identified weaknesses by estimated impact on manuscript quality. Output a numbered revision plan:
## 修订计划(按影响力排序)
1. [HIGH] 标题过于笼统 → 改为包含主要发现的具体标题
当前: "The Role of THBS2 in Cancer"
建议: "THBS2 Overexpression Associates with M2 Macrophage Infiltration and Poor Survival Across 17 Cancer Types"
理由: 具体标题平均被引用量高 22%(Paiva et al., 2012)
2. [HIGH] 摘要缺少定量结论 → 添加关键数字
当前: "THBS2 was significantly upregulated in multiple cancers"
建议: "THBS2 was significantly upregulated in 17/33 TCGA cancer types (Wilcoxon p<0.001), with highest expression in PAAD (HR=2.31, 95%CI: 1.45-3.68)"
理由: 摘要中包含具体数字的论文被引用量高 29%
3. [MEDIUM] Figure 2 caption 缺少统计方法说明
...
Phase 4: Execute Revisions
Apply each revision to the manuscript text. For each change:
- Quote the original text (2-3 lines of context)
- Show the revised text
- State the reason
Constraints:
- Do NOT change any data values, p-values, effect sizes, or sample sizes
- Do NOT add citations that were not in the original or verified through tools
- Do NOT change the interpretation of results
- Do NOT add new claims or conclusions
- DO improve: titles, headings, topic sentences, transitions, figure captions, methods detail, limitation specificity, discussion balance
Phase 5: Re-score and Compare
Re-run ScholarEval on the revised manuscript. Output comparison:
## 审修效果对比
| Dimension | Before | After | Change |
|---------------|--------|-------|--------|
| Novelty | 0.72 | 0.72 | — |
| Rigor | 0.68 | 0.75 | +0.07 |
| Clarity | 0.55 | 0.78 | +0.23 |
| Reproducibility| 0.60 | 0.82 | +0.22 |
| Impact | 0.70 | 0.70 | — |
| Coherence | 0.65 | 0.80 | +0.15 |
| Limitations | 0.40 | 0.75 | +0.35 |
| Ethics | 0.90 | 0.90 | — |
| | | | |
| **Weighted** | **0.66**| **0.76** | **+0.10** |
| **Verdict** | minor_revision | accept | ⬆ |
Citation-impact heuristics: 4/10 → 8/10 factors above threshold
共执行 12 处修订,主要改善了清晰度(+0.23)和可复现性(+0.22)。
核心科学主张和数据未做任何修改。
Phase 6: Output
Save revised manuscript and revision report:
outputs:
📄 reports/manuscript_revised.md — 修订后的完整稿件
📋 reports/revision_report.md — 修订报告(所有变更 + 理由)
📊 reports/scholareval_comparison.md — 评分对比表
Revision Patterns Library
Title Improvements
| Pattern | Before | After |
|---|
| Add key finding | "Role of X in Y" | "X Promotes Y Through Z Mechanism" |
| Add quantitative | "X is associated with Y" | "X Overexpression in N/M Cancers Associates with Poor Survival (HR=...)" |
| Add scope | "Study of X" | "Pan-Cancer Analysis Reveals X as..." |
Abstract Improvements
- Add sample sizes: "patients" → "patients (n=438)"
- Add effect sizes: "significantly different" → "significantly different (Cohen's d=0.82)"
- Add confidence intervals: "HR=2.31" → "HR=2.31 (95%CI: 1.45–3.68)"
Methods Improvements
- Add software versions: "R" → "R 4.3.2"
- Add package versions: "survival package" → "survival package (v3.5-7)"
- Add thresholds: "significant genes" → "genes with |log2FC|>1 and FDR<0.05"
- Add normalization: "normalized data" → "TMM-normalized counts (edgeR v3.40.2)"
Discussion Improvements
- Add counter-argument paragraph: "However, alternative explanations include..."
- Add comparison with conflicting studies: "In contrast to [Author, Year] who found..."
- Convert generic limitations to specific: "small sample size" → "limited sample size in the PAAD cohort (n=178) may reduce power to detect survival differences in subgroup analyses"
Integration with Other Skills
- After any Research Recipe completion, offer: "需要用审修系统优化报告吗?输入 /review"
- Works on any markdown file in the project directory
- Can be applied to user-uploaded manuscripts (paste text or provide file path)