What a prompt is
A "prompt" is just the message you type to the AI. That is the whole definition. You will see the phrase "prompt engineering" and think it is a technical field. It is not. It is writing clearly, in plain English, to a patient helper.
A good prompt is not a magic spell. It is an instruction you would be comfortable giving a capable new colleague on their first day.
If you would describe the task clearly to a person helping you, you already know how to prompt an AI. The techniques below just make the result more reliable, which matters more in research than in everyday use.
The parts of a good prompt
Not every prompt needs all of these, but once you know them you will notice when one is missing.
- Task. What you want done. "Summarise these findings." "Draft a plain-language summary." "Turn this into a table."
- Context. The situation. "This is for a non-specialist community audience." "These are abstracts from a search I have already run." "I am writing the background section of a grant."
- Role (optional). Who the AI should write as. "You are an epidemiologist writing for a ministerial audience." A role sharpens the tone and level.
- Format. What the answer should look like. "A table with these columns." "Under 200 words." "Plain English, no jargon." "Flag anything uncertain."
You do not label these like a form; you write a normal instruction that happens to include them. A useful memory aid from the literature is CLEAR (Lo, 2023): keep prompts Concise, Logical, Explicit, Adaptive and Reflective.
Weak versus strong prompts
Same task, two prompts.
Weak: "Summarise the evidence on diabetes interventions." The AI has no scope, no population, no format. You get something generic and possibly invented.
Strong: "Summarise the evidence from randomised controlled trials published 2020 to 2025 on community health worker-led interventions for Type 2 diabetes in low-income populations. Use only the abstracts I am about to paste. Present a table with columns: Author/Year, Population, Intervention, Comparator, Primary Outcome, Effect Size. Flag anything the abstracts do not state with [VERIFY]."
The strong version names the scope, restricts the AI to material you provide, sets the format, and asks it to mark uncertainty. The difference is usually one or two extra sentences.
Five techniques worth knowing
These come from published guidance for health research (Liu et al., 2025, JMIR; and others) and are the ones that pay off most.
Give it context. The single most useful habit. One or two sentences about your situation, audience and what you have already done turns a textbook answer into one built for your task.
Set a role. "You are a senior epidemiologist writing a policy brief for a non-specialist audience." Assigning an expert persona consistently produces better-pitched output than an unframed request.
Ask it to think step by step. For anything involving reasoning, interpreting a hazard ratio, working through a method, adding "think step by step" improves the quality and reduces errors. This is "chain-of-thought" prompting.
Show an example. If you want a specific format, paste one completed example before asking for the next. For pulling data into a consistent table, one filled-in row dramatically improves consistency. This is "few-shot" prompting.
Build in safety instructions. Frame research prompts with constraints such as: "Only use information in the text I provide. Never invent citations or statistics. Flag any uncertainty with [VERIFY]. If something is not in the text, say you cannot determine it." These noticeably reduce invented content.
A sixth, underused trick: ask the AI to help you write the prompt. "I want to synthesise findings from 20 studies on air pollution and childhood asthma. What should I include in my prompt to get the most useful structured output?" often produces a better final prompt than guessing.
Task-specific notes
For literature synthesis, always paste the actual abstracts or text rather than asking the AI to find papers. This is the single most important habit for reducing invented references. Use dedicated tools to find papers, then an AI to help synthesise what you have verified (see the literature page).
For data interpretation, some tools run real code on data you upload, which is genuine computation, not guessing. Use those for cleaning and charting, and still check the results.
For report and grant writing, specify audience, tone, length and structure, then refine in steps rather than regenerating from scratch.
For search strings, an AI can draft a Boolean string or help structure a PICO question, but the string must be checked by an information specialist before it is used in a formal review.
Keep going, do not restart
The most common beginner mistake is abandoning a prompt because the first answer missed, and starting over. The AI remembers everything in the same conversation and is good at fixing its own work when you say what is wrong.
If the first answer is not right, say what is wrong. The next one is almost always better.
Useful things to say back: "Too long, halve it, same points." "Too formal, rewrite plainly." "The third row is wrong because [why], redo it." "Keep paragraph two, rewrite paragraph one to match its tone."
Spotting when it is wrong
Better prompting does not stop the AI making things up; it makes mistakes easier to spot, because your answer is more specific. Watch for a statistic or quote with no source, very specific claims about small or local topics, and any reference, phone number or web address (these are invented often). A useful closing line for any important prompt: "Before you answer, tell me how confident you are, and where I could verify this." The verification page goes deeper.
A note for this work: do not paste community, cultural or identifying data into a prompt. How to handle research data safely is covered on the responsible-data page, and the data sovereignty and ICIP questions have their own page.
Glossary
- Prompt
- the message you type to an AI.
- Prompt engineering
- the craft of writing prompts that get accurate, useful output; in practice, clear instructions plus a few habits.
- Context
- the background you give the AI so it answers for your situation.
- Role prompting
- telling the AI what expert persona to write as.
- Chain-of-thought
- asking the AI to reason step by step, which improves harder tasks.
- Few-shot prompting
- giving one or more worked examples of the output you want before asking for more.
- CLEAR
- a memory aid for good prompts: Concise, Logical, Explicit, Adaptive, Reflective.
- [VERIFY] flag
- an instruction asking the AI to mark anything it is unsure of, so you can check it.
Take it with you: a one-page cheat sheet of these techniques is available to download from the button at the top of this page.