Paper Submission Review

Paper Submission Review

Pre-submission review of an academic paper draft. Input: a draft (PDF / .tex / extracted .txt). Output: a prioritized list of edits with concrete suggested rewrites and rationales tied to a fixed principle set.

The principles come from comparing real draft → accepted submission diffs and from independent LLM-reviewer cross-checks. The principle reference is at references/principles.md. The expected output structure is at references/report-template.md.

When to use

• The user asks to "review a paper", "check a draft", "do a pre-submission review".
• The user supplies a PDF or .tex file with the goal of polishing it before a venue deadline.
• The user asks "is my paper ready to submit?"

Do not use for:

• Reviewing an already published paper for understanding — use the paper-reader skill instead.
• Reviewing code or notebooks (use review, simplify, or cli-review).
• Generating a new paper outline from scratch.

Input handling

Accept any of: PDF path, .tex path, .txt path, or a directory containing the manuscript.

1. PDF input: extract text with pdftotext -layout <input>.pdf <out>.txt. If pdftotext is missing, install with brew install poppler (macOS) or apt-get install poppler-utils.
2. LaTeX source: read the main .tex and any \input{...} / \include{...} children. Keep section structure visible.
3. Extracted text already on hand: use directly.

For PDFs longer than 20 pages, use pdftotext then read the .txt. Do not page through the PDF visually.

Workflow

Run these phases in order. The principles file (references/principles.md) defines codes P0–P16 (positive principles) and A1–A16 (anti-patterns). Reference them by code in the output.

Phase 1 — Read the whole draft once

Read end-to-end before flagging anything. While reading, capture for yourself (in working memory or scratch notes):

• The paper's title and the noun phrase it uses for its own contribution.
• The first sentence of the abstract; the last sentence of the abstract.
• The headline result(s) from the experiments — the specific numbers and what they say.
• The conclusion's first sentence.
• The section list.

Phase 2 — Run the principle checks

Walk through references/principles.md from P0 → P16 in order. For each, decide whether the draft satisfies it, partially satisfies it, or violates it. Do not skip principles even when they look fine — make the check explicit. Note the anti-patterns (A1–A16) you observe in passing.

The most important checks, in order:

1. P0 Claim–evidence calibration. Compare the verb in the title and abstract ("propose / extend / show / analyze / correlate") against the actual strongest result. If the strongest evidence is a correlation/observation but the paper is framed as a method/application, this is a blocker — flag for retitling and reframing.
2. P1 Named contribution. Find the noun phrase the paper uses for itself. Does it appear consistently in title, abstract, conclusion, and section heading? If absent or inconsistent, flag.
3. P2 Abstract structure. Score the abstract against the background → gap → claim → evidence structure. Flag any abstract that opens with "we propose" or ends with "outperforms baselines" without concrete characterization.
4. P3–P4 Intro and section structure. Underline textbook-level sentences in the intro. Check that section structure mirrors the contribution claim (observation papers need an observation section before the verification section).
5. P5–P7 Operational clarity, reproducibility, metric. For every stochastic / threshold / hyperparameter, check the paper specifies train- and test-time behavior. List missing reproducibility anchors (chosen items, hyperparameter rationale, normalization, tolerance windows, dataset split). For each metric, check whether a naive baseline could game it; if yes, require a justification.
6. P8 Baseline credibility. List the baselines. Flag any missing standard baseline for the field (e.g., DTW for spoken term detection; standard ASR baseline for speech recognition; the obvious published method for the task).
7. P9–P10 Tables and figures. For each table, ask which columns/rows discriminate; flag uniform columns and missing robustness columns. For each claimed comparison, locate the figure and check that comparison is visible in one figure on one page with aligned axes.
8. P11 Interpretation. For every nontrivial number, check there's a "what this means" sentence. For every anomaly (non-monotonic trend, underperforming row), check it's acknowledged or explained.
9. P12–P13 Citations and notation. Check that every standard tool/dataset has a citation; every non-obvious architectural choice has a defending citation; arXiv preprints are upgraded to venue where possible. Flag overloaded symbols (Y as count, etc.) and per-element notation that should be vectorized.
10. P14 Cleanup. Sweep for placeholders (TODO, \cite{?}, XXX, draft emails, missing figure captions).
11. P15 Conclusion alignment. Check the conclusion echoes the abstract's frame. Flag new claims or future-work fishing that the experiments don't support.
12. P16 Length discipline. Identify the lowest-information-density paragraphs as cut candidates if the paper is over its page budget.

Phase 3 — Produce the report

Use references/report-template.md. For each finding:

• Principle: code (P0–P16) or anti-pattern (A1–A16).
• Location: section / page / sentence / table-figure number.
• Quote: the problematic text verbatim, OR a description of what is missing.
• Suggested edit: concrete rewrite, reorder, or addition. Do not be vague ("rephrase this") — provide the actual replacement sentence or the actual missing column header. If you cannot propose a concrete edit, mark this finding "needs author input."
• Rationale: one sentence pointing back to the principle.
• Severity: blocker / important / minor.

Sort by severity, then by principle number (claim-level issues first).

End with a short Summary section: the 3–5 most important moves the author should make.

Phase 4 — Optional cross-check

For high-stakes drafts (e.g. venue with low acceptance rate or the user explicitly asks), dispatch a second independent read via the cli-review skill (Gemini for the prose, Codex for the LaTeX source if available). Append the cross-check findings under "Independent reviewer notes."

Output discipline

• The report is the deliverable. Save it as paper_review.md in the working directory if one exists, otherwise return it in the message.
• Do not modify the draft itself.
• Do not write a "summary of the paper" — the author already knows their paper. The report is only the prioritized edits.
• Concrete suggested rewrites > vague advice. If a principle is satisfied, do not flag it; only flag actionable problems.

Calibration

• A typical 4–6 page conference draft yields 15–35 findings.
• A blocker-class finding (P0 / P1 / P15 mismatch) reframes the paper; expect at most 1–2 of these.
• The majority of findings should be important (concrete editorial moves) — these are what the author will act on.
• minor findings (typo-level, capitalization) are kept terse; group them at the end if there are many.

References

• references/principles.md — full principle catalog (P0–P16 and anti-patterns A1–A16) with worked examples.
• references/report-template.md — the output structure.
• references/examples-v1-vs-v2.md — short worked diffs from the two real draft→accepted pairs the principles are derived from.