Prompt Engineering Is the Skill Nobody Teaches (But Everyone Needs)
Most people use AI like a search engine, typing one vague question and getting one disappointing answer. Better results come from better input, and learning to prompt well improves every task you hand to these tools.
The first time a well-crafted prompt saved me hours, I was preparing a literature review for a grant application in the second half of 2024. I’d been manually reading through dozens of papers, extracting methodological details, and building a comparison table by hand. On a whim, I fed the AI a detailed description of my research question, the specific methodological features I needed to compare, the format I wanted, and three example papers as reference. What came back wasn’t perfect, but it was about 65% of the way there in about two minutes. Cleaning it up took another twenty.
Compare that to my first attempt months earlier: “Help me with my literature review.” The response was a generic, surface-level summary that could have come from a first-year textbook. I closed the tab and went back to doing it manually, convinced the whole thing was overhyped.
I used the same tool both times and changed only what I gave it. Newer models get me 90% of the way there, but the prompting principles haven’t changed.
The Problem: We’re All Prompting Badly
Most people use AI the way they use Google. They type a few words and hit enter. “Summarise this paper.” “Help me with my data.” “Write a literature review.”
Requests like these leave the model guessing at what you want. When the output comes back vague or wrong, people conclude that AI isn’t useful for serious work. I hear this constantly from colleagues: “I tried ChatGPT, it wasn’t very good.” When I ask what they typed, it’s almost always a single sentence with no context.
A search engine retrieves existing documents. An AI model generates a response from everything you give it, so the more you give it, the better the output.
Why Academics Need This Especially
Research tasks are specific. A typical request asks for a comparison of mixed-effects modelling approaches for nested longitudinal data with missing observations. No model can answer that well without knowing your design and how your data went missing.
Think about what you need AI to help with. For a literature review, “find me papers” becomes “compare the methodology across these six studies and identify gaps in how they measure X.” For data analysis, “analyse my data” becomes “I have a repeated-measures design with three conditions, 45 participants, and 12% missing data: suggest an appropriate analysis and write the Python code.” The same goes for grant writing and course material.
The Three Things That Matter
After over a year of using AI assistants for almost everything in my workflow, I give every serious request three things that “help me with my research” lacks: context, materials, and a plan to iterate on.
1. Context
Tell the model who you are, what you’re working on, and who the output is for. With that, it can pitch the response to the right reader and use the methods your field expects. In my honours course, Practical AI for Behavioural Science, I teach students to start every prompt with context, because they immediately see the difference. A prompt that begins “You are helping a fourth-year psychology student analyse qualitative interview data using thematic analysis” produces better output than one that begins “Analyse this data.”
2. Materials
The model can only work with what you give it, in the prompt and in any files you attach. If you’re asking it to review a paper, give it the paper. If you’re asking it to write analysis code, describe the dataset: columns, types, sample size, any quirks. If you’re asking it to draft a rubric, give it the learning objectives, the task description, and an example of a good submission. The real “ghost in the machine” is your understanding of the problem and what solving it requires, and the model sees that understanding only through what you hand over. Most people ask the model to produce something from nothing, then wonder why the result is thin. Hand it your data description, your notes, your old slides, your marking criteria, and the methods sections of relevant papers, and let it work with real material instead of inventing from its training data.
3. Planning and Iteration
Don’t start by asking for the final product. Start by building a plan, with the AI.
People type a prompt, read the response, and either accept it or reject the whole thing. The best results come from working in stages. Describe what you want to achieve, ask the AI to help you structure an approach, review the plan, adjust it, and then ask for the output.
When the output arrives, treat it as a draft. Read it for what’s wrong or missing, then tell the model what to change: “Good, but make the tone more formal.” “The third point is wrong; here’s what I actually mean.” “Now rewrite this for a non-technical audience.” Each round of feedback narrows the output toward what you need. Three or four rounds typically get you to something useful, which one-shot prompting almost never does.
The Trick Most People Miss
Ask the model to help you prompt better. This changed my prompting more than anything else.
Before you ask your actual question, say:
I want to ask you to help me with [task]. Before I do, what information would you need from me to do this really well? Ask me questions.
The model comes back with a list of clarifying questions. Your answers supply the exact context and materials it needs to produce a good response. Modern reasoning models and the agentic systems built on them increasingly do this by default. With models that don’t, the prompt above triggers it.
This is meta-prompting: using the AI to improve your interaction with the AI. It sounds circular, but it works, particularly with models that won’t ask for missing context on their own.
Students Need This Too
Students arrive at university with ChatGPT on their phones and no framework for using it well. They paste in assignment questions and submit whatever comes back, then get mediocre grades and either blame the AI or, worse, conclude that their own thinking doesn’t matter.
I teach prompting as a core skill alongside the domain content, as we do with statistical software and academic databases. Students learn to provide context, feed in materials, iterate on outputs, and verify everything the model produces. Those who learn to prompt well also get better at articulating what they need, which makes them clearer thinkers.
Where to Start
If you want to move beyond vague prompts and start getting useful output from AI assistants, I’ve put together a practical guide with ready-to-use prompt templates for academic workflows: literature review, data analysis, grant writing, course material, and more. It also covers meta-prompting techniques and common mistakes to avoid.
Read the full guide: Prompt Engineering for Academics
I’ll keep the guide current as the models and best practices change.
Michael Richardson Professor, School of Psychological Sciences Faculty of Medicine, Health and Human Sciences Macquarie University
AI Disclosure: I wrote this article with help from AI tools, including Claude. The ideas and the workflow described here are my own. The AI helped me draft and edit the text. I use AI throughout my research and teaching, and I encourage others to do the same. Using AI well is a skill to be learned openly.
The AI models used here, like all current LLMs, were trained on vast quantities of text written by others, largely without explicit consent. Every output they produce draws on the ideas and language of countless researchers and writers. Their labour makes these tools possible, and I want to acknowledge it even where no individual can be credited.