peer-review

name	peer-review
type	skill
description	Peer review of research funding applications and academic submissions. Scheme-agnostic — fetches current criteria from the relevant handbook each round, since weights and language change. Covers Detailed Assessor and College-of-Experts / General Assessor roles, plus collegial draft review.
category	instruction
triggers	["review this grant","review this application","future fellowship","DECRA","ARC assessment","ERC review","SNSF review","fellowship review","detailed assessor","general assessor","college of experts","process reviewer comments"]
modifies_files	true
needs_task	true
mode	workflow
domain	["academic"]
allowed-tools	Read,Write,Edit,Grep,Glob,Bash
version	3.0.0
permalink	skill-peer-review

Peer Review Skill

Prepare peer reviews of research funding applications or academic submissions, matching the current round's scheme criteria and producing evidence-based, signable feedback. The agent prepares and verifies; the academic owns the scores, the net call, and the final submitted text.

What this skill is, in one paragraph

A review is not a linear pipeline that ends in a self-check. It is an adaptive loop of five stages whose quality comes from two moves a naive draft-then-submit flow lacks: adversarial independent verification of every claim, and a voice-match + whole-set de-template pass that makes the prose signable and un-formulaic. Its failures happen at the seams between stages — a polish pass that quietly promotes a paraphrase into a fabricated quote; a voice rewrite that drops a verified correction. So the loop gates its regeneration boundaries: "verified" attaches to a committed artifact, not an idea, and any regeneration re-enters verification.

Core Rules

• Evidence-Based: Every claim must cite the application (page, section, line, or quote).
• Absent Evidence Is Feedback: A missing required element is a scoreable weakness, not an omission to skip silently — but say where you looked (a documented grep with synonyms), never assert an absence you didn't search for.
• Fetch Round Criteria: Fetch the current scheme guidelines (ARC, ERC, SNSF, NHMRC) for this round. Weights, bands, and character minimums change; do not reuse prior criteria.
• Human Owns Judgment: Scores, top-line read, net call, and submitted text are the academic's. The agent prepares evidence-mapped drafts with scores left blank. This is how funding-body GenAI policies (e.g. ARC's ban on AI-generated assessor text) are satisfied structurally — not by refusing to help draft.
• No Integrity Allegations in Comments: Route suspected research-integrity breaches to the scheme's integrity office separately (see [[platform-instructions]]).
• One Living Draft in Git: One draft per review, versioned by git history — no -v2 / -v3 files, no carried-forward appendix. Cut critique lives in the history, not in a growing reserve block (see [[review-template]]).
• Every Worked Example Verbatim-True: Any example in these skill files that quotes an application must be a real, grep-checked string. A fabricated "illustrative" quote in the doctrine is how the doctrine teaches the bug. Generic placeholders () are fine; invented specifics are not.