# Review Paper > Adversarial paper review simulating a skeptical referee — checks identification, statistical claims, robustness, and presentation against real referee patterns - Skill: `thinkingwithagents/review-paper` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add thinkingwithagents/review-paper` - Raw SKILL.md: https://api.skillmd.com/api/skills/thinkingwithagents/review-paper/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: thinkingwithagents (https://skillmd.com/u/thinkingwithagents) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/thinkingwithagents/review-paper --- # Review Paper *v1.1 — Pre-submission self-review combining econ-native evaluation with adversarial stress testing* Simulate a rigorous peer review of an academic manuscript before submission. Combines economics-specific review criteria (identification strategy, econometric specification) with a Devil's Advocate stress test and editorial synthesis. **Argument:** `$ARGUMENTS` - Path to a paper file (`.pdf`, `.tex`, `.qmd`, `.md`, `.docx`) - Or a project name (will look for the paper in the project's standard locations) **Modes** (append to argument): - `full` (default) — 3-pass review: Econ Referee + Devil's Advocate + Editorial Synthesis - `quick` — Single-pass referee report only (skip DA and synthesis) - `methodology` — Deep dive on identification and econometrics only - `da-only` — Devil's Advocate stress test only (skip balanced review) Example: `/review-paper ~/Dropbox/Github/bd-social-norms/paper/main.tex full` Example: `/review-paper dep-dts quick` --- ## Instructions ### Step 0: Locate and Read the Paper 1. If `$ARGUMENTS` contains a file path, read that file directly 2. If `$ARGUMENTS` is a project name, check these locations in order: - `~/Dropbox/Github/[project]/paper/` - `~/Dropbox/Github/[project]/manuscript/` - `~/Dropbox/Github/[project]/draft/` - Glob for `*.tex`, `*.pdf`, `*.qmd`, `*.docx` in the project repo 3. For PDFs, read in chunks (5 pages at a time). For LaTeX, read the main file and any `\input{}` files. 4. Read the **full paper** end-to-end before starting any review pass. Parse the mode from `$ARGUMENTS`. Default to `full` if not specified. 5. **Load the reference file** `review-paper-references/real-referee-patterns.md` via the Skill tool. This contains calibration patterns from real referee reports at top journals — use it to ground the review in how actual referees write and what they prioritize. ### Step 1: Econ Referee Pass Think like a referee at a top-5 economics journal (AER, QJE, Econometrica, JPE, REStud). Evaluate across these 6 dimensions: #### 1.1 Argument Structure - Is the research question clearly stated and well-motivated? - Does the introduction establish the question's importance in 2-3 paragraphs? - Is the logical flow sound (question -> literature -> method -> results -> conclusion)? - Are conclusions supported by the evidence presented? - Are limitations acknowledged honestly? #### 1.2 Identification Strategy - Is the causal claim credible? What is the source of exogenous variation? - Are the key identifying assumptions stated explicitly? - Threats to identification: omitted variables, reverse causality, measurement error, selection, spillovers, SUTVA violations? - For RCTs: randomization quality, compliance, attrition, balance, pre-analysis plan adherence? - For quasi-experimental: parallel trends, exclusion restriction, first-stage strength, bandwidth sensitivity? - Is the estimator appropriate for the research design (TWFE concerns with staggered adoption, etc.)? - **Implementation fidelity** (for program evaluations): Was the program delivered as designed? Do deviations (delays, mistargeting, short duration) limit what the paper can claim? Is the paper evaluating the program-as-designed or program-as-implemented? #### 1.3 Econometric Specification - Standard errors: clustered at the right level? Robust? Few-cluster corrections needed? - Functional form: appropriate? Sensitivity to alternatives? - Sample selection: any concerns about conditioning on post-treatment variables? - Multiple testing: how many hypotheses tested? Any corrections? - Are point estimates economically meaningful (magnitude, not just statistical significance)? - Heterogeneity analysis: pre-specified or exploratory? Appropriate corrections? - Power: is the study adequately powered for the effects being detected (or not detected)? For null results, report MDEs — what effect sizes can be ruled out? - **Mechanism evidence**: Does the paper provide *positive* evidence for the claimed mechanism, or only rule out alternatives? Suggest specific heterogeneity analyses that would discriminate between competing explanations. - **Cost-benefit** (for interventions): Is return on investment discussed? Even back-of-envelope cost per unit of outcome is valuable. #### 1.4 Literature Positioning & Framing - Are the key papers in this area cited? - Is prior work characterized accurately and fairly? - Is the contribution clearly differentiated from existing work? - Any missing citations a referee would flag? - Is the paper positioned in the right literature (not just the convenient one)? - **Framing breadth**: Could the paper be framed more broadly? Narrow framing undersells the contribution. Avoid "first paper to..." claims — they're debatable and shrink the scope. Help the author see the biggest defensible question their paper answers. - Does the intro explain what the results *mean* for the existing literature, not just that the paper "contributes to" it? #### 1.5 Writing Quality - Clarity and concision (economics papers should be tight) - Abstract effectively summarizes question, method, findings, and contribution - Consistent notation throughout - Tables and figures are self-contained (clear labels, notes, sources, sample sizes) - Appropriate length for the contribution #### 1.6 Presentation & Data Transparency - Are summary statistics presented and discussed? - Balance tables (for RCTs)? Do they show treatment and control means separately (not just the difference)? Randomization inference? Joint test? - Are variable definitions clear and documented? For composite indices, are the components listed? - Is the data section informative about sample construction? - Are results presented clearly (coefficient, SE, significance, N)? - Are all tables/figures referenced in the text? - **Admin vs. self-reported data**: If both exist, are both shown? Are discrepancies discussed as a data quality signal? - **Table parsimony**: Can ITT and IV be consolidated (one in main tables, one in appendix)? Can related tables be combined? - **Self-reported outcomes**: Are table titles explicit when outcomes are self-reported? - **Robustness to controls**: Is there an appendix showing results without baseline controls? - **Displacement/GE effects**: For labor market papers, is this discussed alongside main results (not just in the conclusion)? Rate each dimension 1-5 (1 = major problems, 5 = excellent). Generate **3-5 referee objections** — the tough questions a top referee would ask. For each: - State the objection clearly - Explain why it could be fatal or damaging - Suggest how the author could address it **If mode = `quick` or `methodology`**: Write the review report (see output format below) and stop here. For `methodology`, focus only on dimensions 1.2 and 1.3. --- ### Step 2: Devil's Advocate Stress Test *Skip if mode = `quick` or `methodology`.* Switch perspective entirely. You are no longer a balanced reviewer — you are an adversarial critic whose job is to find every vulnerability. This is separate from Step 1; do not repeat the same points. Run these 8 challenges: #### 2.1 Core Thesis Challenge Construct the **strongest possible argument against** the paper's main conclusion. Write 200-300 words making the case that the paper is wrong. #### 2.2 Cherry-Picking Detection - Are results selectively reported? Are there specifications that were likely run but not shown? - Does the paper emphasize its strongest results while burying weaker ones? - Is the sample definition suspiciously convenient for the results? #### 2.3 Confirmation Bias Scan - Does the paper only cite evidence supporting its hypothesis? - Are contradictory findings from other studies discussed? - Is the theoretical framework chosen because it predicts the "right" result? #### 2.4 Logic Chain Validation Trace the argument from assumptions through to conclusions. Identify any point where the chain breaks or requires an unstated leap. #### 2.5 Overgeneralization Check - Do the results generalize beyond the specific sample/context studied? - Does the paper claim more generality than the evidence supports? - External validity: would this hold in other countries/time periods/populations? #### 2.6 Alternative Explanations List 2-4 alternative explanations for the main results that are **not adequately ruled out**. For each, explain what additional evidence would distinguish the author's story from the alternative. #### 2.7 Stakeholder Blind Spots - Whose perspective is missing from the analysis? - Are there distributional effects or welfare implications not discussed? - Would practitioners/policymakers interpret these results differently than the authors intend? #### 2.8 "So What?" Test - If the results are exactly right, what changes? Who acts differently? - Is the contribution incremental or transformative? - Does the paper answer a question people are actually asking? Classify each finding as: - **CRITICAL**: Fatal flaw in core argument — cannot be rescued without fundamental revision - **MAJOR**: Seriously undermines credibility but can be addressed with additional analysis/writing - **MINOR**: Worth noting but does not affect core argument - **OBSERVATION**: Not a defect, but an alternative perspective worth considering **If mode = `da-only`**: Write the Devil's Advocate report and stop here. --- ### Step 3: Editorial Synthesis *Only runs in `full` mode.* Now step back and synthesize Steps 1 and 2 into an editorial decision. Think like an editor making a desk decision. #### 3.1 Consensus Analysis - Where do the referee evaluation and Devil's Advocate agree? (These are the strongest signals.) - Where do they disagree? (The DA may have been too harsh, or the referee too generous.) - Which DA findings are genuine threats vs. hypothetical concerns unlikely to bother a real reviewer? #### 3.2 Decision Based on the synthesis: - **Strong Accept**: Excellent across all dimensions, DA found only minor/observation-level issues - **Accept with Minor Revision**: Strong paper, 1-2 addressable issues, no DA critical findings - **Revise and Resubmit**: Good potential, but 2+ major issues or 1 DA critical finding that can be addressed - **Reject**: Fundamental identification/argument problems that cannot be fixed with revision #### 3.3 Revision Roadmap Organize all findings into a prioritized action list: - **Priority 1 (Required)**: Must fix before submission — fatal or near-fatal issues - **Priority 2 (Strongly Recommended)**: Will significantly strengthen the paper - **Priority 3 (Nice to Have)**: Polish items, additional robustness checks, framing improvements --- ### Step 4: Write the Report Save the full report to the working directory as `review_[sanitized_paper_name]_[YYYY-MM-DD].md`. Tell the user the full path to the output file. --- ## Output Format ```markdown # Manuscript Review: [Paper Title] **Date:** [YYYY-MM-DD] **Mode:** [full / quick / methodology / da-only] **File reviewed:** [path to manuscript] **Reviewer:** /review-paper skill v1.0 --- ## Summary Assessment **Overall recommendation:** [Strong Accept / Accept with Minor Revision / Revise and Resubmit / Reject] [2-3 paragraph summary: main contribution, key strengths, and central concerns. Be direct.] --- ## Econ Referee Report ### Dimension Scores | Dimension | Score (1-5) | Key Issue | |-----------|:-----------:|-----------| | Argument Structure | | | | Identification Strategy | | | | Econometric Specification | | | | Literature Positioning | | | | Writing Quality | | | | Presentation & Data | | | | **Overall** | **[avg]** | | ### Strengths 1. [Strength — be specific, cite sections/tables] 2. [...] 3. [...] ### Major Concerns #### MC1: [Title] - **Dimension:** [which of the 6] - **Issue:** [specific description] - **Suggestion:** [how to address it] - **Location:** [section/page/table] [Repeat for each major concern] ### Minor Concerns #### mc1: [Title] - **Issue:** [description] - **Suggestion:** [fix] [Repeat] ### Referee Objections #### RO1: [The tough question] **Why it matters:** [why this could be fatal or damaging] **How to address it:** [suggested response or additional analysis] [Repeat for 3-5 objections] --- ## Devil's Advocate Report *(Only in `full` and `da-only` modes)* ### Strongest Counter-Argument [200-300 words. The single best case that the paper's conclusion is wrong.] ### Critical Findings [Any CRITICAL-severity issues. If none, state "No critical findings."] ### Major Findings [MAJOR-severity issues with challenge number (2.1-2.8)] ### Minor Findings & Observations [Brief list] ### Alternative Explanations Not Ruled Out 1. [Alternative — what evidence would distinguish it] 2. [...] --- ## Editorial Synthesis *(Only in `full` mode)* ### Consensus Points [Where referee and DA agree — these are the most important signals] ### Areas of Disagreement [Where DA was harsher than the referee assessment warrants, or vice versa] ### Decision Rationale [2-3 sentences explaining the recommendation] ### Revision Roadmap #### Priority 1: Required Before Submission - [ ] [Action item with specific guidance] - [ ] [...] #### Priority 2: Strongly Recommended - [ ] [Action item] - [ ] [...] #### Priority 3: Nice to Have - [ ] [Action item] - [ ] [...] --- ## Specific Comments [Section-by-section or line-by-line comments, if any. Include page/paragraph references.] ``` --- ## Principles - **Be constructive.** Every criticism comes with a suggestion. The goal is to make the paper better, not to tear it down. - **Be specific.** Reference exact sections, equations, tables, page numbers. - **Think like a top-5 referee.** What would make them recommend rejection? What would make them enthusiastic? - **Distinguish fatal flaws from polish.** Not everything is equally important. Prioritize. - **Acknowledge what's done well.** Good research deserves recognition. Acknowledge difficulty of execution: "Successful evaluations of government programs are very challenging." - **Economics-specific lens.** Identification, not just "methodology." Economic significance, not just statistical significance. Credibility revolution standards. - **Be honest about uncertainty.** If you're unsure whether data exists, say so: "I don't know if this data exists, but..." - **Offer choices, not mandates.** For presentational suggestions, give the author latitude: "I would leave it to you to decide." - **Framing advice is a gift.** When a paper is framed too narrowly, suggesting a broader frame helps the author — it's not a criticism. - **Do NOT fabricate.** If you cannot read a section clearly, say so. Do not invent details about the paper. - **Do NOT hallucinate citations.** If you suggest missing references, flag that the user should verify they exist.