10 Claude prompts I use weekly for paper writing

Why Prompt Structure Matters More Than the Model

The single biggest lesson from using LLMs on real manuscripts is that an unstructured prompt produces fluent, confident, and frequently wrong output. Models are very good at reproducing the shape of scientific writing — the cadence of a Methods section, the hedging of a Discussion — without any commitment to whether the underlying claims are true. Left open-ended, they will happily invent a trial name, a statistic, or a citation to fill a gap. The prompts below exist to remove that freedom: each one constrains the model to organize facts I supply rather than generate facts of its own.

When I sit down to write, I don't want magic. I want reliability. I've tested dozens of tools, and most fail when subjected to the rigors of peer review. I use my own workflows to ensure that the data I present is accurate, verifiable, and free of hallucinations.

Why I Use Structured Prompts

I spent far too long tweaking prompts to get a usable output. I finally sat down and codified the exact prompts I use for clinical research writing, and I keep them in a dedicated Claude Project so they are one paste away on every paper. The payoff is less time staring at a blank page and far fewer factual errors to catch later. I found that strict constraints are the only way to get clinical relevance out of an LLM.

The IMRaD Constraints

I always break my prompts down by section. I never ask the model to write an entire paper. I prompt specifically for the Introduction, Methods, Results, and Discussion. I provide the exact data points. I explicitly forbid the model from hallucinating specific trial names, authors, or DOIs. I learned this lesson the hard way, which I wrote about in /blog/citation-hallucination-ai-writing.

The 10 Prompts

Every prompt below follows the same rule: I supply the facts, the model organizes them. Paste your own material where the brackets are. Each one ends with an instruction to flag gaps instead of filling them, because a visible gap costs me two minutes and an invented fact can cost me the paper.

1. Introduction from a verified gap

You are helping me draft the Introduction of a clinical research paper.
Here are the only sources you may use, each with its key finding:
[1] ... [2] ... [3] ...
Our research question: [question]
Write a 3-paragraph Introduction: what is known, what is not known, what this study does.
Cite only the numbered sources above using [n]. Do not add any study, author, statistic or guideline that I did not provide.
Where an argument needs support I have not given you, write [CITATION NEEDED] instead of inventing one.

2. Methods from protocol notes

Below are my protocol notes and data dictionary. Convert them into a Methods section following [STROBE / CONSORT / CARE].
Use only details present in my notes. Do not assume sample size calculations, software versions, ethics approval numbers or statistical tests.
After the draft, list every reporting-guideline item that my notes do not cover, as questions for me.
Notes: [paste]

3. Results from a table, without interpretation

Describe the results in the table below in plain prose for the Results section.
Report numbers exactly as they appear. Do not round, recompute, or add p-values, confidence intervals or effect sizes that are not in the table.
Do not explain why results occurred; interpretation belongs in the Discussion.
If the table is ambiguous (unclear denominators, missing units), stop and ask.
Table: [paste]

4. Discussion skeleton

Using only the Results and the sources below, outline a Discussion with:
1) the principal finding in one sentence,
2) how it compares with each provided source (agree, disagree, or not comparable, and why),
3) limitations that follow from the Methods I gave you,
4) what the finding changes, if anything, for practice or research.
Mark any sentence that goes beyond the supplied material with [UNSUPPORTED].
Results: [paste]  Sources: [paste]

5. Abstract under a word limit

Write a structured abstract (Background, Methods, Results, Conclusions) of at most [250] words from the manuscript below.
Every number in the abstract must appear in the main text; after the abstract, list each number with the section where it appears.
The conclusion must not claim more than the Results show.
Manuscript: [paste]

6. Structural edit that keeps every fact

Edit the draft below for logical flow and academic register.
You may reorder sentences, merge or split paragraphs, and tighten wording.
You may not change, add or remove any number, citation, drug name, dose or outcome.
Return the edited text, then a change log listing each structural change and why.
Draft: [paste]

7. Unsupported-claim audit

Go through the text below sentence by sentence.
List every sentence that makes a factual or causal claim without a citation or a reference to our own data.
For each, say what kind of evidence would be needed to support it. Do not suggest specific papers.
Text: [paste]

8. Hostile Reviewer 2 on the methods

Act as a demanding peer reviewer for [journal]. Review only the Methods and Results below.
Give major and minor concerns: design threats, confounding, selection bias, missing data handling, mismatch between stated and reported outcomes, and reporting-guideline gaps.
No praise, no summary of the paper. Rank the concerns by how likely they are to cause rejection.
Methods and Results: [paste]

9. Response-to-reviewers letter

Draft a point-by-point response to the reviewer comments below.
For each comment: quote it, give a courteous response, and state exactly where in the revised manuscript the change was made, using the change list I provide.
If a comment is not addressed in my change list, write [AUTHOR DECISION NEEDED]. Do not promise analyses or changes that are not on my list.
Comments: [paste]  Changes made: [paste]

10. One defensible novelty sentence for the cover letter

From the Results and the gap statement below, propose three one-sentence novelty claims for a cover letter to [journal].
Each claim must be supported by our Results alone. For each, state what an editor could check in the manuscript to verify it, and what existing evidence would make the claim false.
Do not use "first study", "to the best of our knowledge" or "fills an important gap".
Results: [paste]  Gap: [paste]

I explain the reasoning behind that last one in Cover Letter Drafting with AI.

How I Use Them

I keep these prompts in one Claude Project together with the study protocol and the final tables, so the model always has the source material in context. I run one prompt per section and never chain them into a single "write my paper" request. Every output gets checked against the source tables before it goes into the manuscript. For a broader look at which model handles which writing task best, see Claude vs GPT-5 vs Gemini for paper writing; for stress-testing a finished draft, see red-team prompts for weak claims.

My Favorite Prompt

One of my most reliable prompts is for structural editing. I feed the model my raw, messy draft and ask it to enforce a specific logical flow without changing any facts. I tell it to flag any claims that lack a citation. I demand that it maintain my academic register. I use a variation of this prompt to critique my methodology, asking the model to act as a hostile Reviewer 2. I have packaged these and more advanced workflows into my Paper Structuring Prompt Pack, available at gumroad.com.

Connecting the Workflow

I use these prompts in conjunction with my overall tool stack, which I outline in /blog/ai-research-stack-5-tools-that-save-time. I also combine this with my Zotero workflow, detailed in /blog/zotero-claude-literature-synthesis. I treat the LLM as a Junior Analyst who is eager to please but prone to making things up. I manage it with strict instructions and rigid verification.