Peer Review

Peer Review

Evaluate an academic manuscript the way a careful journal peer reviewer would: find the real weaknesses, judge the strengths honestly, and turn that into (1) a structured written review and (2) simulated scores on nine dimensions (1–5, where 5 is best). Grounded in official reviewer guidance from Wiley and Cambridge University Press (see references/reviewing-guide.md).

Two modes — confirm which one

This skill serves two audiences. Figure out which from context; ask only if genuinely unclear.

• Author self-check (作者自检) — the user wrote the paper and wants to know

what a reviewer would criticize before submitting. Tone: a candid but supportive mentor. Goal: maximize the chance of acceptance by fixing problems now. Be direct about flaws — sugar-coating helps no one — but always pair each weakness with a concrete fix.

• Reviewer report (替我审稿) — the user was invited to review this manuscript

and wants help drafting their reviewer report and recommendation. Tone: the professional, constructive voice of a reviewer writing to the editor and author. Same analysis; the output is framed as a report rather than advice.

The underlying evaluation is identical. Only the framing of the output differs (see "Output" below).

In reviewer-report mode, if the user is still deciding whether to accept the invitation, help them check the three gates first — expertise, conflict of interest, and time — and note the journal's review type (single/double-blind, open). See "Before you accept an invitation" in references/reviewing-guide.md. Once they've accepted (or for self-check mode), proceed to Step 1.

Step 1 — Read the manuscript

Get the full text first. Manuscripts usually arrive as files:

• .docx → use the docx skill to extract text (preserve headings, tables,

figure captions).

• .pdf → use the pdf skill to extract text. If it is scanned/image-only,

OCR it.

• Pasted text or other formats → work with what's given.

If a file is password-protected or unreadable, say so and ask for an unlocked version rather than guessing at the content.

Note the target journal if known — relevance and scope judgments depend on it. If unknown, ask, or evaluate against general standards for the field and say you've done so.

Step 2 — First pass (form an overall impression)

Skim the whole paper once before judging details, mirroring how reviewers work. Form a view on the big questions and watch for major flaws that can sink a paper regardless of polish:

• Conclusions that contradict or outrun the evidence; unsupported claims.
• Outdated, inappropriate, or unreliable methodology.
• Key processes / prior work / confounders overlooked.
• Insufficient, contradictory, or cherry-picked data.
• Problems with figures, tables, or images (incl. signs of manipulation).

A genuine major flaw caps the recommendation no matter how strong the rest is — note these explicitly.

Step 3 — Detailed section-by-section pass

Go through the manuscript section by section. Use the checklist in references/reviewing-guide.md ("Detailed reading — what to evaluate") so nothing is missed — Introduction, Methods, Results/Discussion, Conclusions, Figures/Tables, References, plus an originality/plagiarism sanity check. Take notes tied to section + page/line numbers so every comment is locatable and actionable.

Step 4 — Score the nine dimensions

First, fix the venue type — it changes the scoring bar. Conference papers and journal articles are not held to the same standard, so confirm which one this is:

• If the user said (or it's obvious from the file/venue) whether it's a

conference paper or a journal article, use that.

• Otherwise ask once: 会议论文 conference / 期刊论文 journal?

The nine dimensions and the 1–5 scale stay the same; what shifts is the bar for each score and how strictly you weight things. In short: journals expect full rigor, completeness, and a substantial contribution; conference papers are typically shorter and faster-turnaround, so reward novelty/timeliness and tolerate narrower scope, fewer baselines, and "early but promising" work — while still requiring the core method to be sound. The "Venue calibration" section in references/scoring-rubric.md spells out the difference; read it and state which bar you applied.

Score each dimension 1–5 (5 = best) using the anchored rubric in references/scoring-rubric.md. Read that file before scoring — each dimension has explicit descriptors for what a 1, 3, and 5 look like, so scores are calibrated rather than arbitrary.

The nine dimensions:

| Dimension | In one line | |---|---| | Interest | How compelling is the question to the journal's readers? | | Originality | How novel are the question / approach / data / findings? | | Contribution | How much does it advance theory, practice, or knowledge? | | Relevance | How well does it fit the journal's scope and field? | | Theory | How sound is the theoretical grounding / framework? | | Methodology | How rigorous and appropriate is the research design & analysis? | | Presentation | How clear is the writing, structure, and exhibits? | | Validity | Do the evidence and analysis actually support the claims? | | References | How accurate, current, and balanced is the literature? |

For every dimension, give the score plus a one- to three-sentence justification (why this score, not higher or lower) plus a concrete improvement the author could make to raise it. A bare number is not useful — the reasoning and the fix are the point.

Overall recommendation

After scoring, give a holistic recommendation — do not just average the numbers. Weight validity and methodology heavily, and remember a single fatal flaw caps the outcome. Map to the standard editorial options:

• Accept — publishable largely as is.
• Minor revision — sound; fixable issues, no new analyses needed.
• Major revision — promising but needs substantial work (often new

analyses/data) and re-review.

• Reject — fatal flaws, or contribution/fit too weak to salvage here.

State which one and why in one short paragraph.

Output

Output language is the user's choice — confirm it. Before writing the review, decide the language:

• If the user stated a preference (e.g. "用中文" / "in English" / "中英对照"),

honor it.

• If the manuscript and request are clearly one language and the user gave no

preference, default to that language and say so in one line.

• If it's genuinely ambiguous, ask once: 中文 / English / 中英对照(bilingual)?

Then write the entire review — section headings, scores table, and all comments — in the chosen language. Use the bilingual headings below only if the user picked 中英对照; otherwise translate the headings into the single chosen language (e.g. English-only: "Summary / Scores / Major issues / Minor issues / Recommendation"). Keep the structure identical regardless of language:

## 总体评价 / Summary
[2–4 sentences: what the paper does, its core contribution, and your overall
take. Lead with a genuine strength before the criticism.]

## 评分 / Scores (1–5, 5 = best)
| 维度 Dimension | 分数 Score | 理由 Why | 改进建议 How to improve |
|---|---|---|---|
| Interest | x/5 | ... | ... |
| Originality | x/5 | ... | ... |
| Contribution | x/5 | ... | ... |
| Relevance | x/5 | ... | ... |
| Theory | x/5 | ... | ... |
| Methodology | x/5 | ... | ... |
| Presentation | x/5 | ... | ... |
| Validity | x/5 | ... | ... |
| References | x/5 | ... | ... |

## 主要问题 / Major issues
[Numbered. The things that most threaten publication. For each: what it is,
where (section/page/line), why it matters, and what would fix it.]

## 次要问题 / Minor issues
[Numbered. Clarifications, wording, citation gaps, mislabeled exhibits, small
factual/numerical errors.]

## 推荐意见 / Recommendation
[Accept / Minor revision / Major revision / Reject — with a one-paragraph
rationale.]

Mode-specific framing:

• Author self-check → phrase major/minor issues as a prioritized to-do list

("先修复这些再投"); make the recommendation a prediction ("以现状投稿，审稿人 大概率会给 Major revision，因为……").

• Reviewer report → phrase as comments addressed to the author/editor; keep

any sensitive concerns (suspected plagiarism, ethics, fraud, conflicts) separate as "给编辑的保密意见 / Confidential comments to the editor".

How to write the comments — tone and ethics

These come straight from the official reviewer guidance; full detail in references/reviewing-guide.md.

• Be constructive. The point of a review is to help the work improve, even

when recommending rejection. Critique the research, never the authors.

• Be specific and locatable. Number points; cite section/page/line. "Unclear

on p.7" beats "parts are unclear".

• Be honest but fair. Don't soften a real flaw into invisibility, and don't

inflate scores to be nice — a misleadingly rosy self-check fails the author at submission. Lead with strengths, then problems.

• Judge content over language. For non-native English writing, flag only

errors that obscure meaning; don't copyedit grammar line by line.

• Stay in scope and objective. Evaluate against the journal's aims and the

field's standards, not your personal preferences or what you would have done.

• Ethics flags go to the editor. Suspected plagiarism, data fabrication,

undisclosed conflicts, duplicate publication, or unethical procedures belong in confidential editor comments, written as if the authors might read them.

Reference files

• references/scoring-rubric.md — the 1–5 anchored descriptors for all nine

dimensions. Read before scoring.

• references/reviewing-guide.md — distilled Wiley + Cambridge guidance: review

types, accepting an invitation, the section-by-section checklist, major-flaw list, how to structure a review, tone, ethics, and the four recommendation routes. Read before the detailed pass.

Source & license

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

• Author: lyra81604
• Source: lyra81604/peer-review
• License: MIT

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