Writing WorkflowMay 18, 2026

Self-Peer-Review with AI: A 5-Step Manuscript Audit Workflow

Before you submit your manuscript to a high-impact journal, you owe it to yourself (and your co-authors) to catch the "stupid" errors first. I built this 5-step audit after a reviewer caught a statistical discrepancy in a pilot draft that I should have spotted myself. Since then, I've refined this workflow into a systematic self-peer-review process using AI that now serves as the backbone of the AVR Paper Checker on aiforacademic.world.

Here is the 5-step workflow to audit your manuscript before the editor even sees it.

Step 1: Claim Mapping (The Logic Audit)

The first step is to ensure that your manuscript actually says what you think it says. After writing the same paper for weeks or months, we often get so close to the text that we miss gaps in logic — claims that float without evidence, and data that never gets properly connected to a conclusion.

The Prompt Strategy:

Feed your Introduction and Discussion to an LLM and ask it to extract the primary claims and the evidence provided for each. The prompt I use:

"Read this manuscript section. For each claim in the Discussion, identify: (1) the specific result from the Results section that supports it, and (2) the table or figure number where that result appears. Flag any claim where the supporting evidence is absent, indirect, or requires an inference not stated in the text."

If the AI can't find a clear link between a claim and a specific result, neither can a reviewer. What you'll get back is a table of claims, their supporting data, and — most usefully — a list of "unsupported claims." Every item on that list is a rewrite target.

Common failure modes this catches:

Discussion conclusions that go beyond what the Results actually show ("our findings suggest" followed by a statement the data doesn't support)
Introduction claims cited from secondary sources when you have primary data that disagrees
Passive constructions like "it has been shown that..." with no citation

Spend 20 minutes here. The worst thing a reviewer can say is "the conclusions are not supported by the data" — and this step makes that remark nearly impossible.

Step 2: Statistical Sanity (The "Did I Forget a Test?" Check)

I once missed a heterogeneity diagnostic in an early meta-analysis draft — Claude flagged it in 30 seconds when I asked it to audit my Methods against my Results. The assumption had been buried so long in my mental model of the analysis that I had stopped seeing it.

The Prompt Strategy:

Provide your Methods and Results sections together. Use this prompt:

"You are a statistical reviewer for a medical journal. Read the Methods section carefully. For each statistical test described, list: (1) the formal assumptions required for that test to be valid, (2) whether the Results section explicitly states that those assumptions were tested, and (3) any missing assumption checks. Flag each gap as HIGH, MEDIUM, or LOW severity based on how likely a reviewer is to raise it."

This gives you a checklist of assumption checks with severity ratings. HIGH severity items (e.g., violation of normality for a parametric test, violation of the proportional hazards assumption for a Cox model) need to be addressed before submission — either by running the test or by adding a limitation statement.

What this catches in practice:

Parametric tests applied without normality checks in small samples
Repeated-measures ANOVA without sphericity tests
Binary logistic regression without collinearity checks
Meta-analyses reporting heterogeneity (I²) without sensitivity analysis when I² > 50%

Run this on every version of your manuscript, not just the final one. Statistical errors introduced during revision are the ones that most often slip through.

Step 3: Reference-Claim Alignment (The Citation Audit)

Reviewers trust real DOIs, but they are increasingly wary of "citation laundering" — where a real paper is cited for a claim it doesn't actually support. This is not always intentional: you cite a review article that cites the original study, and by the time the claim appears in your manuscript it has drifted from what the primary source actually showed.

This is a known failure mode in AI-assisted writing, too. I've written at length about AI citation hallucination and how to catch it before submission. The short version: never trust a citation you haven't personally verified against the abstract.

The Workflow:

Use CiteCheck (our open-source tool at pypi.org/project/citecheck) to verify that all DOIs in your reference list resolve to real, published papers.
For your 10 most critical citations — those supporting the strongest claims in your Discussion — use this prompt:

"Here is the abstract of [paper title]. Here is the sentence in my manuscript that cites it: '[sentence]'. Does the abstract support this citation as written? If not, explain the gap and suggest how the sentence should be revised to accurately reflect what the paper shows."

If the paper doesn't support the claim, revise the sentence before submission — not after a reviewer flags it.

A citation audit on 10 papers takes about 15 minutes. It has saved my manuscript from at least two "citation does not support claim" reviewer comments. At top-tier journals, those comments can result in desk rejection.

Step 4: Novelty Defense (The "So What?" Test)

Editors at top-tier journals (BMJ, NEJM, Lancet) decide on desk-reject in under a minute based on the perceived novelty of the contribution. The fastest way to fail this test is to describe what you found without making clear why it matters in a way that no prior paper has achieved.

The Prompt Strategy:

"You are a cynical senior editor at a high-impact medical journal. Your job is to desk-reject papers that lack genuine novelty. Read my Abstract and Discussion. Generate three specific, evidence-based reasons why my findings are NOT novel — reasons a real editor might use to justify rejection. Be direct. Do not soften the critique."

Use these three "attacks" to strengthen the Novelty and Significance paragraph in your Discussion. For each attack the AI generates, write one counter-argument with specific evidence from your results. If you can't counter it, that's a real weakness — and it's better to discover that now than after a rejection.

I've run this on every manuscript I've submitted in the past year. My most recent paper went back to the writing stage twice after Step 4. That felt frustrating. The paper that came out the other side got accepted without major revision.

Step 5: Reviewer Red-Teaming (The Hostile Review)

The final step is the most brutal: the hostile review. This is where you stop asking the AI to help and start asking it to attack.

The Prompt Strategy:

"You are Reviewer 2. You are skeptical, methodologically strict, and you have read 500 papers this year in this area. Your job is to find every reason to reject this manuscript. Read the full paper (Methods + Results + Discussion) and give me: (1) the three biggest methodological weaknesses, (2) the three most over-reaching claims in the Discussion, and (3) two questions about missing data or missing analyses that you would require the authors to address in revision. Be specific — cite section numbers and line numbers where possible."

The output is a mock review. Read it as if it came from a real reviewer. For every point, decide: is this a real weakness I need to address, or is this a misread I can rebut in a response letter? Both answers are useful. The first group needs to go back into the manuscript. The second group becomes your revision letter.

Running this step before submission means you arrive at peer review with a pre-prepared response to the most predictable criticisms. You know where the weak spots are. You've already thought through the rebuttals.

How to Chain All Five Steps

Run the steps in this order for maximum efficiency:

Claim Mapping first — fix logic gaps before you check citations, or you may be defending claims you'll cut anyway.
Statistical Sanity second — statistical errors often cascade into Discussion claims, so fix these before the novelty audit.
Reference-Claim Alignment third — with the logic and stats clean, now verify the citations that support the surviving claims.
Novelty Defense fourth — now that the content is solid, evaluate whether the contribution is actually clear.
Reviewer Red-Teaming last — the final stress test on the finished product.

Each step takes 20–40 minutes. Total time for a full audit: 2–3 hours. Compare that to a rejection cycle: 3–6 months lost plus revision work.

A Note on What AI Audits Cannot Do

This workflow is not about letting AI review your science. It's about using AI to make your human-led research defensible. The AI does not know whether your findings are true. It does not know your field's standards well enough to replace a domain expert. What it does well: systematic checks against stated criteria, pattern recognition in text, and generating adversarial critiques on demand.

Use it for those three things. Do the actual scientific judgment yourself.

By the time you finish this 5-step audit, your manuscript will be significantly more defensible. For teams who run this workflow on every submission, we've productized the full pipeline in the AVR Paper Checker at aiforacademic.world/avr. If you want to skip the manual prompting and get a structured audit report in minutes, that's the fastest path — but even the manual workflow above will raise your submission quality substantially.

The five steps, one more time:

Claim Mapping — does every claim trace to evidence?
Statistical Sanity — were all required tests run?
Reference-Claim Alignment — do your citations actually say what you say they say?
Novelty Defense — can you rebut the "so what?" attack?
Reviewer Red-Teaming — have you already thought through Reviewer 2's objections?

Run it before every submission. It will save you more time than it costs.