Oleafly Review Manuscript

Review a manuscript

Produce the review a careful referee would write: specific, anchored to the text, and honest about what you could not check. You read and you report. You do not edit.

The one rule

Never modify a manuscript file. No write_file, replace_in_file, create_file, rename_file, or delete_file on anything that is part of the document. The only file you create is the report under review/. If the user asks you to apply a fix during the review, stop, finish the report, and offer to make the edits as a separate task.

Before you start

Ask, or state your assumption, if any of these is unclear:

• Is this the user's own draft, or someone else's submission they are refereeing?
• Which venue, and is the review single-blind, double-blind, or open?
• What should the review focus on?

For an unpublished submission belonging to someone else, keep everything local. Do not put manuscript sentences into literature_search, alphaxiv_search, verify_citation, or any other network tool. Those queries leave the machine. Reading, compiling, and searching inside the project do not.

Procedure

1. Map the project

Call list_files for the file tree, then project_map for the structure: sections, labels, bibKeys, macros, theorems, inputGraph, unresolvedRefs, unresolvedCites. Record the sections you will read with update_todos so the pass is visible and nothing is skipped.

If project_map is not offered (the engine does not advertise a document index), fall back to search_project for \section, \subsection, or # headings and build the outline by hand.

2. Read the full source

read_file every source file, following the include order from inputGraph. read_file truncates at 800 lines or 40,000 characters per call, so keep calling it with a higher offset until you reach the end of each file. Read the bibliography file too.

Do not review from the abstract and the section headings. A referee report written without reading the methods is worthless, and it shows.

3. Read the rendered document

Call compile, then get_pdf_text to see what the reader actually sees: float placement, table overflow, figure ordering, page breaks in the wrong place, and anything the source hides.

If the compile fails, call get_log, put "the manuscript does not compile" at the top of the report with the first error and its location, and review the source alone.

4. Load the checklists

Call load_skill with peer-review, then read_skill_file for the parts you need:

• references/common_issues.md for recurring problem patterns and constructive phrasings
• references/statistical_reproducibility.md for methods and statistics review
• references/reporting_standards.md when a reporting guideline applies

Call load_skill with scholar-evaluation and read_skill_file for references/evaluation_framework.md to get the five scoring dimensions and the rules that go with them.

Reporting guidelines apply only to the study types they were written for. Use CONSORT for randomized trials, STROBE for observational epidemiology, PRISMA for systematic reviews and meta-analyses, ARRIVE for animal experiments. For a theory paper, a systems paper, or a machine-learning methods paper none of them applies. Say so in one line rather than forcing a checklist that does not fit.

If those two skills are not enabled, or read_skill_file is not available in this run, use references/referee-checklist.md and references/scoring-rubric.md in this skill instead. They carry the same shape at less depth.

5. Anchor every remark

For each issue, find the exact sentence with search_project and write down the file and the line number. Quote the text you are objecting to, trimmed to the part that matters.

A remark you cannot anchor does not go in the report. "The statistics are weak" helps nobody. "methods.tex:112 reports a t-test on counts with n=7 per group and no normality check" can be acted on.

6. Write the report

Create review/report.md with create_file, then fill it with write_file. Follow assets/report-template.md. The sections are:

1. Summary. Three to six sentences: what the paper claims, how it argues it, and your overall read. No accept or reject verdict. That belongs to an editor.
2. Strengths. What genuinely works, with locations. Skip this and the report reads as hostile and gets discounted.
3. Major comments. Issues that affect whether a central claim stands. Numbered. Each one carries location, observation, why it matters, and the specific action that would resolve it.
4. Minor comments. Issues worth fixing that do not threaten a claim.
5. Line notes. The table below.
6. Scoring. Per dimension, from references/scoring-rubric.md or the scholar-evaluation framework.
7. Limits of this review. What you could not assess: unavailable data, methods outside your competence, a failed compile, sections that were empty.

The line notes table:

Severity is major, minor, or note. Sort the table by file, then by line.