peer_review_assistant v1.0.0

name: "peer_review_assistant" description: "Helps an agent produce structured academic peer reviews and calibrated calibration practice. Triggers when asked to review a manuscript, evaluate claims and methods, map reporting guidelines, write major and minor comments, or draft a point by point author response letter." includeInPrompt: true

Peer Review Assistant

Purpose

Turn a manuscript and its context into a clear, consistent, venue compliant review you could send to an editor, plus a calibrated stance on recommendation. The skill separates claim mapping, method adequacy, reporting guideline fit, reproducibility, and language calibration so agents do not skip steps or invent facts.

Workflow

1. Intake and safety.

• Identify paper type, scope, and journal or venue policy.
• Confirm you can keep the manuscript confidential, and do not upload the full text to external generative AI tools. If you used any AI help in evaluating claims, note it only where the venue requires it.
• Note conflicts and blind status.

1. Claim map.

• List the central claims in your own words. For each claim, attach the exact evidence the manuscript provides and any missing link. If evidence is absent, say so.

1. Methods evaluation.

• Check design fit for the stated goal.
• Assess sample, controls, randomization, blinding, measurement, statistical plan, pre registration if applicable, processing code and data access.
• Flag reproducibility gaps. Be specific about what is missing, not why it feels off.

1. Reporting guideline mapping.

• Match design to guideline by EQUATOR principles: CONSORT for randomized trials, PRISMA for systematic reviews or meta analysis, STROBE for observational studies, ARRIVE 2.0 for animal work, STARD for diagnostics, CARE for case reports, SRQR or COREQ for qualitative work, CHEERS for economic evaluations, TRIPOD for prediction models.
• Pull the current checklist from the guideline home. Do not quote long lists from memory.

1. Scoring and structure.

• Write a short summary statement with the research synopsis, overall recommendation, two to three strengths, two to three weaknesses, and a bottom line on significance and soundness.
• Major comments are numbered, critical, validity focused, each with a clear issue, why it matters, and a specific fix or experiment.
• Minor comments are numbered, focus on clarity, presentation, and completeness.
• Recommendation must be consistent with the major comments. If not, fix one or the other.

1. Calibration practice.

• Use the severity guide in references/calibration.md. Keep language proportional: needs clearer reporting, would benefit from, limited to, alternative explanation. Avoid harsh phrasing when a gentler fit is accurate.
• Run the practice exercises in that file with the provided fictional manuscripts.

1. Author response practice.

• Draft a point by point reply. Quote each comment, answer under it, mark where you changed the manuscript with exact sections, figures, or tables. Never skip a point. If a request is outside scope, explain why.

Output Contract

• Input: manuscript metadata, paper type, venue policy link if known, and the text or a section excerpt you are allowed to use.
• Output: markdown with sections: Summary, Major comments, Minor comments, Recommendation, Reproducibility checklist, and Response letter draft if requested.
• If you cannot check a detail, place it in Unverified gaps and do not assert it as fact.

Operating Rules

1. Verify every factual claim against a source you can cite. Unverifiable statements get cut or reframed as opinion.
2. Do not invent numbers, dates, quotes, or events. If a number matters and you do not have it, do not make one up.
3. Keep manuscripts confidential. Do not upload the manuscript to external AI tools. If your workflow used AI assistance, follow venue disclosure policy.
4. Align recommendation and evidence. A reject recommendation without fatal major issues, or an accept with unresolved fatal issues, fails this rule.
5. Be specific about locations. Reference sections, figures, tables, page or line numbers where available.
6. Respect blind review. Do not reveal your identity or signal who you are when the venue requires anonymity.
7. Style. Plain, direct sentences. No em dashes. No promotional language. No AI cliches.