How to Write Good AI Prompts for Scientific Work: A 5-Part Template

In short: A good prompt tells the model who you are and what the situation is, gives one clear task, names the reader, specifies the output format and states how you'll judge the result, including what it must not do. Missing pieces are filled by guesswork. Below: the template, three before-and-after examples from scientific work and the mistakes to avoid.

Why your prompt gets a generic answer

You paste an abstract and type "summarize this." What comes back is accurate, bland and no use for the decision you actually need to make: should anyone on your team read the full paper? The model did what you asked. It just couldn't know who you are, why you're asking or what a useful answer looks like.

Specific prompts get specific answers. Five parts cover most scientific tasks.

The five parts

1. Context: who you are and the situation

One or two sentences. "I'm a postdoc in tumor immunology preparing a journal club" gives the model a level, a field and a purpose.

2. Task: one verb, one deliverable

"Summarize", "compare", "critique", "rewrite", "extract". Combining five requests in one message produces five half-answers. Do them in sequence.

3. Audience and tone

Who will read the output: a reviewer, a clinician, a program officer, a business-development lead who is not a scientist? If you have a paragraph whose style you like, paste it as a sample.

4. Format and length

A table with named columns, 150 words, three options, bullets with one sentence each. If you don't specify, the model chooses, and it usually chooses too long.

5. Checks: what good looks like and what is forbidden

The most commonly missing part, and the most valuable in science. Examples: "use only the text I provide", "quote the sentence that supports each claim", "say 'not reported' instead of guessing", "flag any number you are unsure about", "ask me questions before you start if something is missing".

A template to copy

Context: I'm a [role] in [field], working on [project or goal]. [One or two sentences of situation.]
Task: [one verb] [what exactly].
Audience and tone: for [reader], in a [neutral / formal / plain-language] tone. Style sample: [optional paragraph].
Format: [table with columns... / N words / N options].
Checks: use only the material I paste. Support each claim with a quote from it. Write "not reported" where the text is silent. Flag anything you are unsure about. If you need more information, ask me up to three questions first.

[Paste source text here]

Three before-and-after examples

Example 1: Reading a paper for a decision

Before:

Summarize this paper.

After:

I'm an R&D lead in a biotech company deciding whether our team should invest time in the approach in the paper below. I have a biology background but limited statistics training.
Summarize it in four parts: research question, method, main result, limitations the authors state themselves. Then add a fifth part: "What I would want to know before relying on this", based only on the text.
Plain language, under 250 words in total. Quote the sentence behind the main result. If the sample size is small or the result comes from a single experiment, say so explicitly. Do not add information that is not in the text.

[paste paper text]

Example 2: Drafting a methods paragraph

Before:

Write a methods section for my experiment.

After:

I'm preparing the methods section of a manuscript for [journal type]. Below are my lab notes on the protocol.
Write one methods paragraph in the usual past-tense, third-person style, using only the details in my notes. Do not add reagents, concentrations, catalog numbers, software versions or durations that I did not state.
After the paragraph, list every detail a reviewer would expect that is missing from my notes, as a checklist for me to fill in.

[paste notes]

Example 3: Answering a reviewer comment

Before:

Help me respond to this reviewer.

After:

I'm the corresponding author responding to Reviewer 2. Their comment is below, followed by what we actually did or can do (my notes).
Draft a polite, direct response of up to 150 words that (1) acknowledges the point, (2) states what we changed and where in the manuscript, and (3) if we are not making the change, explains why without defensiveness.
Use only what is in my notes. Mark any place where you would need a page number or a new result from me as [NEED].

Comment: [paste]
My notes: [paste]

Mistakes that cost the most in research

1. Letting the model supply facts. Without "use only the text I provide", it fills gaps from general knowledge, which includes plausible but wrong details.
2. Trusting numbers and citations. Check each figure and reference against the source.
3. Pasting sensitive material. Unpublished data, identifiable patient information and partner-confidential content do not belong in unapproved tools. See Can Pharma and Biotech Teams Use ChatGPT and Claude Safely?.
4. Accepting the first draft. The first output is a starting point. Iterate in the same conversation.
5. Asking for interpretation. The model can summarize and organize your evidence. What a result means for your hypothesis is your call.

Follow-up prompts that improve a result

• "What in this answer is least supported by the text? List it."
• "Rewrite it for someone outside the field, without losing any limitation."
• "Which of these claims would a skeptical reviewer challenge first, and why?"
• "Shorten to 60% of the length and keep every number."

Next steps

Once your prompts are reliable, put the stable parts in a standing instruction file so you don't retype them. Claude Code for Researchers shows how, and AI Literature Review Without Fake Citations applies the template to a full workflow.