Who is Manuscript review for?
Reviewers holding a manuscript that is not theirs to share, with a report due.
GuidesPlaybook
Manuscript review
A structured report: a summary the authors recognise, major issues with fixes, minor issues, citation flags with public URLs, and a recommendation that follows from the majors.
IntermediateSteps5
Time~110 min
Tools5 tools
Before you start
- The manuscript and the venue, editor, or course policy
- Enough domain knowledge to judge the method, not only the writing
Tools in this guide
0%0/5 steps
Peer reviewer · 90–120 minutes
The manuscript in your inbox is unpublished and confidential. Paste it into a chatbot and the review is compromised before you have written a line of it.
This runbook settles the venue rule first, then builds the report from one human read plus the public citation record.
Structured referee report
Structured referee report
A summary in the authors' own terms, major issues with a suggested fix each, minor issues, a citation flag table with public URLs, and a recommendation with one sentence of reasoning.
Major 2: section 4 reads the coefficient as causal, but the design is a single cross-section with no instrument. Fix: recast as an association, or add the panel the authors mention in the limitations. Recommendation: major revision.
Open the journal, conference, or course policy on generative AI in peer review and copy the sentence that covers submitted work. Write one word at the top of your notes: allowed, forbidden, or unclear. Unclear counts as forbidden.
A one-line policy lock with the quoted sentence and the URL you read it on.
Find the official page where [venue] states its policy on generative AI in peer review. Return the URL and nothing else. Do not summarise the policy.
Pitfall
Taking the summary instead of the page. Publishers revise these policies mid-cycle, and the version a model learned is last year's.Verification
You can quote the sentence and name the page it sits on.Read the manuscript straight through with the pen down. Then write the main claim in two sentences from memory. File the entry in Zotero with venue, deadline, confidentiality status, and the policy lock. The PDF stays on your machine.
Two sentences in your own words, plus a Zotero entry carrying the handling note.
Pitfall
Starting margin comments on page two. You spend the rest of the review arguing with sentences instead of with the design.Verification
You can state the main claim without opening the PDF again.Build a four-column table: claim, evidence offered, method behind it, conclusion drawn. Mark the row where the chain breaks. If the lock says allowed, Claude can pressure-test the map you wrote. If it says forbidden or unclear, the map stays in your notes and nothing leaves the machine.
A claim, evidence, and method table with the weakest link marked and a reason beside it.
Only if the venue permits unpublished material. Here is my own claim map, in my words: [notes]. Which step is least supported by the evidence I recorded? Do not add methods or literature I did not name.
Pitfall
Confusing “I would have done it differently” with “the evidence does not support this”. The first belongs in minor comments, if anywhere.Verification
You can name the weakest link and the passage in the manuscript that creates it.Work through the load-bearing references. In Scite, check whether later papers cite each one as supporting or contrasting. In Semantic Scholar, check that the paper exists as described and that the novelty claim survives a search. This step touches published work only, so it runs even when the lock says forbidden.
A citation flag table: DOI, what the manuscript says it shows, what the record shows, public URL.
For this DOI: [doi], summarise how later papers cite it, supporting versus contrasting, with the strongest example of each. I will open the records myself. Do not comment on any unpublished work.
Pitfall
Writing “citation needed” from a list you never opened. Some of those references exist and say close to the opposite of what you assumed.Verification
Every flag links to a public record a co-reviewer can open.Open with a summary in the authors' own terms so they can see you read it. Then major issues, minor issues, and a recommendation that follows from the majors. Every major comment names the problem, says why it changes the conclusion, and offers one thing the authors could do.
The structured referee report, ready for the submission system.
Here is my issue list, with no manuscript text in it: [issues]. Which comments state a problem without a possible fix, and does my recommendation match the number and severity of the majors? Do not soften the wording.
Pitfall
Recommending rejection because the minor list got long. Count the majors, then pick the recommendation.Verification
Each major comment carries an action, and the recommendation follows from the majors.You should now have: Structured referee report
- The policy lock is written and quoted before any cloud tool sees manuscript text.
- The two-sentence summary came out of your read, not out of a model.
- Every citation flag has a public URL, every major issue has an action, and the recommendation matches the majors.
PrivacyUnpublished manuscripts are confidential. Where the venue, editor, or course forbids generative AI on submissions, the PDF does not go into a cloud tool, and neither do excerpts, paraphrases, or an anonymised version. Scite and Semantic Scholar only touch published papers, which is why the citation work runs either way. Treat an unclear policy as a ban and say so in your reviewer notes.
Questions this playbook answers
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What do you finish Manuscript review with?
A summary in the authors' own terms, major issues with a suggested fix each, minor issues, a citation flag table with public URLs, and a recommendation with one sentence of reasoning.
How long does Manuscript review take?
90–120 minutes
What goes wrong in Manuscript review?
Each step names the mistake you are about to make:
- Taking the summary instead of the page. Publishers revise these policies mid-cycle, and the version a model learned is last year's.
- Starting margin comments on page two. You spend the rest of the review arguing with sentences instead of with the design.
- Confusing “I would have done it differently” with “the evidence does not support this”. The first belongs in minor comments, if anywhere.
- Writing “citation needed” from a list you never opened. Some of those references exist and say close to the opposite of what you assumed.
- Recommending rejection because the minor list got long. Count the majors, then pick the recommendation.