A simple framework for structuring prompts that consistently produces useful output.

CONTEXT: What does the AI need to know about your situation?

• Your field/domain (biology, psychology, economics)
• Your research question or problem
• Your data characteristics (200 samples, 50 variables, longitudinal)
• Your constraints (limited sample size, observational data, budget constraints)
• Assumptions or background the AI should know

TASK: What exactly do you want the AI to do?

• Analyze, explain, summarize, generate, compare, identify, suggest
• Be specific about the output you want
• Single task usually works better than "do this and also this and also this"

FORMAT: What form should the output take?

• Bulleted list, table, paragraph, code, outline, narrative
• Length (brief, detailed, comprehensive)
• Audience (for my advisor, for a journal, for a collaborator, for myself)
• Any specific structure (e.g., "organized by methodology," "grouped by outcome")

Example without framework:

• Prompt: "Help me with my literature review"
• Result: Long generic response not tailored to your needs
• Time wasted: 15 minutes reading irrelevant output

Example with framework:

• Prompt: "I'm writing a literature review on cognitive behavioral therapy for anxiety in adolescents (psychology). My review will focus on methodological approaches used in randomized controlled trials. I've identified 15 core papers. TASK: Organize the main methodological approaches these papers use. FORMAT: Create a table with columns for [Author Year], [Sample Size], [Anxiety Measure], [Treatment Duration], [Control Condition]. I'll use this to identify patterns in how the field approaches measurement and comparison."
• Result: Exactly the table you need for your review
• Time saved: Hours vs. starting from scratch

More specific prompts produce more useful outputs than general prompts.

Key points:

• Vague: "Analyze this data"
• Specific: "I have fMRI data from 30 participants (15 anxiety, 15 control). TASK: Generate code for whole-brain voxel-wise comparison using FSL, including thresholding at p < 0.05 FWE-corrected. FORMAT: Provide bash script with comments explaining each step."

• Vague: "Help me think about my research design"
• Specific: "I want to test whether cognitive training (8 weeks, 2x/week, online format) improves reading comprehension in 9-year-olds with dyslexia. I have access to 50 children (25 intervention, 25 waitlist control) from three schools. TASK: Identify potential confounds I should measure or control for. FORMAT: List 5-7 key confounds organized by category (child-level, school-level, family-level)."

• Vague: "How do I write this section?"
• Specific: "I have a methods section describing an intervention but it's unclear how exactly the 'dose' of the intervention varied between participants. My advisor said the current version 'sounds like everyone got the same thing' but actually dose varied. TASK: Revise to clarify how dose varied. FORMAT: Show me alternative wording that makes the variation clear while keeping it concise (the entire methods must stay under 2,000 words)."

Your first prompt won't be perfect; use iteration to get better outputs.

Key points:

• First prompt: Exploratory, to see what the AI understands
• "That's closer, but I need to focus on X instead of Y"
• Second prompt: Refined based on first response
• "Now I need this organized differently"
• Third prompt: Further refinement
• Good researchers rarely get perfect output first; they refine through iteration
• Each iteration costs minimal time; the first prompt teaches both you and the AI more about what you need

Example iteration:

1. First: "Generate code for analyzing treatment effect"
2. AI generates basic analysis code
3. You: "This looks good but I need to account for the hierarchical structure (students in schools). Can you revise?"
4. AI generates mixed-effects version
5. You: "Perfect. Can you also add code to check assumptions?"
6. AI adds assumption-checking code
7. Result: You have exactly what you need through iteration

Different tasks need different prompting approaches.

For literature synthesis:

• CONTEXT: What papers have you read? What topic are you synthesizing?
• TASK: What should the AI synthesize? (Key findings? Methodological similarities? Contradictions?)
• FORMAT: Organization structure? Briefing document vs. detailed analysis?
• Example: "I've read 8 papers on neurofeedback for ADHD. TASK: Identify what these papers agree on regarding mechanism of action. FORMAT: Organize by proposed mechanism, noting which papers support each."

For research design feedback:

• CONTEXT: What are you trying to study? What constraints do you have? (Sample size, budget, ethical considerations)
• TASK: What aspect do you want feedback on? (Feasibility? Validity? Ethical issues? Statistical power?)
• FORMAT: How detailed should the response be? (Checklist? Narrative discussion? Specific suggestions?)
• Example: "I want to study whether anxiety decreases after yoga intervention. I can recruit 20 participants from community yoga studios. I only have funding for pre-post measurement. TASK: Identify threats to internal validity. FORMAT: List each threat and briefly explain why it's a problem for my design."

For code generation:

• CONTEXT: What language? What are you analyzing? (Data structure, sample size, outcome)
• TASK: What analysis do you want to run? (Specific analysis type, not "analyze my data")
• FORMAT: Should the code include comments? Comments level of detail? Include assumption checking?
• Example: "I have data in CSV format with 300 observations and variables: age, depressionscore, treatment (yes/no), timepoint (baseline, 1-month, 3-month). TASK: Generate R code for linear mixed-effects model predicting depressionscore by treatment, timepoint, and their interaction, with random intercept for participant. FORMAT: Include comments explaining each line and code to check model assumptions."

For writing improvement:

• CONTEXT: What's the current draft doing? (Explaining a concept? Presenting results? Summarizing literature?) What's your audience?
• TASK: What specific improvement do you want? (Clarity? Conciseness? Tone? Grammar?)
• FORMAT: Show me suggestions inline? Provide a revised version? Explain why the revision is better?
• Example: "This paragraph explains why social media use affects sleep (my results). My advisor said it reads like I'm certain about the mechanism but I haven't tested it. TASK: Revise to acknowledge uncertainty about mechanism while presenting the evidence. FORMAT: Show the revised paragraph and explain what changed and why."

Providing examples helps AI understand what you want; sometimes it's better to describe what you want.

Provide examples when:

• You want a specific format or structure
• The task is hard to describe but easy to show
• You have an example of good output

Example prompting:

• "Here's an example of a table I like: [show table]. Create a similar table for my data, with columns for [X, Y, Z]"

Guide instead of example when:

• Your example might bias the AI toward something you don't actually want
• The task is easy to describe
• You want originality rather than copying a style

Descriptive prompting:

• "Create a table with columns for author, year, sample size, and main finding. Keep it to 2-3 sentences per finding, focusing on surprising or important results."

Practical Research Use Cases

Scenario: You're writing the background section of a paper. You've identified 12 key papers on your topic. You want AI to help organize your synthesis.

Poor prompt: "Summarize these papers"

Better prompt using CONTEXT-TASK-FORMAT:

"I've read 12 papers on microbiota's role in anxiety disorders. The papers come from different fields: some from neurobiology journals (focusing on mechanism), some from psychology journals (focusing on clinical outcomes), some from microbiota journals (focusing on what causes microbiota changes). TASK: Show me how these papers cluster together thematically. Are there consistent themes across all papers, or do different fields focus on different aspects? FORMAT: Create an outline organizing the papers by theme, noting where different fields agree or disagree. I'll use this structure for my literature review."

Why it works:

• CONTEXT: AI knows you have 12 papers, understands the interdisciplinary nature
• TASK: Clear ask—cluster by theme, identify agreement/disagreement
• FORMAT: You'll get an outline structure you can directly use

Scenario: You want to generate Python code for exploratory data analysis of your dataset.

Poor prompt: "Write Python code to explore my data"

Better prompt using CONTEXT-TASK-FORMAT:

"I have a CSV file with 500 observations and 45 variables: demographics (age, sex, education), clinical measures (depression, anxiety, social functioning), and biomarkers (cortisol, inflammatory markers). Missing data is minimal (< 5%). TASK: Generate Python code for exploratory data analysis including: distribution plots for each continuous variable, correlation matrix visualization, and descriptive statistics by group (depression: yes/no). FORMAT: Use pandas and matplotlib; include comments for each section; include code to check for and display missing values."

Why it works:

• CONTEXT: Clear about data structure, size, content
• TASK: Specific analyses listed
• FORMAT: Tool preferences, comment level, specific checks desired

Scenario: You're designing a study and want to think through validity threats.

Poor prompt: "Is my study design good?"

Better prompt using CONTEXT-TASK-FORMAT:

"I'm designing a quasi-experimental study comparing anxiety outcomes in patients who receive standard anxiety treatment vs. standard treatment plus technology coaching. Participants will be assigned based on what my clinic can offer each month (not random). I'll measure anxiety at baseline and 3 months. I can recruit 60 participants. TASK: Identify the biggest threats to internal validity in this design. FORMAT: For each threat, explain: (1) Why it's a threat, (2) How likely is it to affect my results, (3) What could I do to reduce it? Focus on practical solutions given my constraints (no randomization possible, 3-month follow-up)."

Why it works:

• CONTEXT: Clear about study type, assignment method, measurement points, sample size, constraints
• TASK: Specific—identify threats, not "is it good?"
• FORMAT: Structured response you can use for your proposal

Scenario: You have a methods section that's technically accurate but hard to follow.

Poor prompt: "Make this better"

Better prompt using CONTEXT-TASK-FORMAT:

"This is my methods section describing our fMRI analysis pipeline. It's technically accurate but a reviewer said it was 'hard to follow the sequence of steps.' TASK: Improve clarity about the sequence of analysis steps. Make it easier for readers to understand what happens first, second, third. FORMAT: Revise just enough to improve clarity without adding length. I'll use this in a journal paper, so it should match scientific writing standards. Show me the revised version and point out the specific changes you made and why."

Why it works:

• CONTEXT: Understanding the problem (technical but unclear)
• TASK: Specific problem to solve (sequence clarity)
• FORMAT: Revision with explanation so you can learn from it

Hands-On Exercise

Objective: Learn to write effective prompts through iteration.

Steps:

1. Choose a research task you want to try with AI (literature organization, code generation, feedback on a design, writing improvement)

1. Write three versions of a prompt for this task:

• Version 1: Write what you might naturally ask (probably vague)
• Version 2: Rewrite using CONTEXT-TASK-FORMAT structure
• Version 3: Add more specificity based on thinking about Version 2

1. Try Version 1 prompt with ChatGPT, Claude, or your chosen tool:

• Record the output
• Note: Is this usable? Missing important details? Too generic?

1. Try Version 2 prompt with the same tool:

• Record the output
• Compare to Version 1: Did it improve? How?
• Is the output more focused on your specific needs?

1. Try Version 3 prompt with the same tool:

• Record the output
• Did adding more specificity help further?
• At what point did you get diminishing returns?

1. Analyze the differences:

• What specific changes made the biggest difference in output quality?
• What information, if you'd left it out, would have made output much worse?
• What detail didn't matter much?

1. Document your learning:

• Write down 3-5 principles you learned about prompting
• Create a template for this type of task that you'll use in future research

Time required: 60-90 minutes

Common Mistakes and Misconceptions

AI is very literal. "Help me with my data" could mean a hundred things. AI will pick one and start there, often wrong. Specificity is not excessive; it's necessary.

Iteration is part of the process. Your first prompt teaches you and the AI what you need. Refining the prompt based on the first response produces better output than perfectionism on the first try.

Sometimes. But irrelevant context confuses AI. Include information the AI needs to answer your question, not everything about your research.

It's not. Providing an example of a good table format is smart prompting. Use examples to show format or style; don't rely on examples to guide analysis.

Format matters. "Organize as a table with columns X, Y, Z" produces a table. "Explain the main findings" produces paragraphs. Be specific about format to get the structure you need.

Key Takeaways

• The CONTEXT-TASK-FORMAT framework structures effective prompts: CONTEXT provides necessary background, TASK specifies exactly what you want the AI to do, FORMAT describes what form the output should take
• Specificity matters: Detailed prompts focused on your actual research situation produce focused, useful output; vague prompts produce generic, often unhelpful output
• Iterative refinement is normal: First prompt teaches you and the AI what you need; refine based on output to get better results
• Different tasks need different prompt structures: Literature synthesis, code generation, research design feedback, and writing improvement each have their own CONTEXT-TASK-FORMAT patterns
• Examples help when you want a specific style or format; description works when you want originality or are asking for analysis
• Good prompts are reusable: Once you craft a good prompt for a task type, you'll use similar versions repeatedly, refining them as your research evolves

Reflection Questions

1. Your typical research questions: Think about a research task you need help with. Write out a vague prompt you might naturally ask, then rewrite it using CONTEXT-TASK-FORMAT. How different are they?

1. Prompt library: For your research, what 3-4 types of prompts would you use regularly? (literature organization, code generation, design feedback, etc.) Write templates for each.

1. Iteration strategy: For a complex research task, how would you approach iterative refinement? What would you ask in the first prompt to learn what you need?

1. Specificity challenges: What aspects of your research are hardest to describe to someone (or an AI) who doesn't know your field? How would you provide sufficient context without overwhelming detail?