EPPS Math & Coding Camp
AI Prompt
Engineering
Writing prompts that actually give the model what it needs.
JP AdjadehEPPS, UT Dallas
Before We Start
Which AI tool do you actually use?
Show of hands.
ChatGPT
Claude
Gemini
Comet AI
Copilot
Others
I do not use AI at all
A Fair Question
Can we trust AI?
Every major AI tool tells you the same thing, right on its own screen.
ChatGPT
ChatGPT can make mistakes. Check important info.
Claude
Claude can make mistakes. Please double-check responses.
Gemini
Gemini can make mistakes, so double-check it.
Comet AI
Comet can make mistakes. Verify important information.
Copilot
Copilot uses AI. Check for mistakes.
Perplexity
Perplexity can make mistakes. Check important info.
The Core Problem
Hallucination
- Fluent, confident answers can still be wrong
- Fabricated citations, invented numbers, plausible-looking errors
- No built-in signal that anything is off
Tell us about a time AI gave you a wrong answer — one that, once you fact-checked it, it admitted was wrong.
The Discipline That Resolves This
Human in the loop
Every issue just named traces back to one thing: the model generates, but nothing checks it unless you do.
The One Rule
Never run or trust AI output you have not personally checked. If you can't verify it, don't rely on it yet.
- Verify numbers by hand — recompute a statistic or spot-check a few rows
- Verify code by tracing it — running without an error is not the same as correct
Example 1
What's wrong with this prompt?
I need to understand American democracy for a paper I'm writing.
- No role — no lens to respond through
- No task — "understand" isn't a deliverable
- No context — "a paper" says nothing about the argument or angle
- No format — length, tone, structure all unset
Same Ask, Rebuilt
Here's the same request, done well
Weak
I need to understand American democracy for a paper I'm writing.
Strong
You are a PhD student studying comparative political institutions. Explain how democratic accountability functions in the United States, focusing on the checks between the executive and legislative branches, and note one theoretical debate about how effective this mechanism actually is. Write this as a single analytical paragraph for a doctoral seminar audience, under 150 words.
The weak prompt sounds like it has context — "a paper" — but tells the model nothing it can actually use.
The Framework
Five things the model can't guess on its own
R
Role
Who should the model act as?
T
Task
One specific, concrete ask
C
Context
What it has no way of knowing
F
Format
Bullets, table, length, tone
E
Examples
A model output to imitate
R–T–C–F–E. Every strong prompt today follows this shape — including the ones you'll write for R code this afternoon.
Same Problem, Your Field
Pick the card closest to your program
Each of these has the same gaps as Example 1 — "my data," "my research," "my paper" are gestured at but never described.
GIS"I have some GIS data on land use — what should I do with it?"
Criminology"I'm looking at crime data for my research — what does the data say?"
Economics"I found this economic trend in my data — what does it mean?"
Political Science"I need some background on political institutions for my paper."
Public Affairs & Policy"Can you help me with a policy memo / look at some policy data?"
Social Data Analytics"I have some social media data — what can I do with it?"
Int'l Political Economy"I need background on trade for my paper."
Next: rebuild the one that's yours using R–T–C–F–E. Real topic, real dataset if you have one.
Exercise 1
Run it weak, then run it strong
- 1Run your field's weak prompt in your AI tool of choice. Look at what comes back.
- 2Rebuild it using R–T–C–F–E, swapping in your own real topic or dataset.
- 3Run the rebuilt version. Compare the two outputs side by side.
- 4Note which single addition — role, task, context, format — changed the output the most.
Live Share
Show us your before / after
One volunteer per program. Project your screen and walk us through:
- 1The weak version you started with, and what it got wrong
- 2Your rebuilt version
- 3The one change that mattered most
Check Your Rebuild
Model answers, by field
"You are a PhD student using GIS to study land-use change. I have parcel-level land-use data for a mid-sized U.S. county, 2010–2020. Suggest three exploratory spatial analyses (e.g., hot-spot mapping, density surfaces) that could reveal patterns of urban sprawl, and explain what each would show methodologically. Bullet points, under 200 words."
"You are a PhD student in criminology analyzing city-level crime data. My dataset has monthly reported crime counts by offense type for one city, 2015–2023. Suggest three visualizations appropriate for detecting seasonal versus long-term trends, and note one limitation of each. Bullet points, under 200 words."
"You are a PhD student in economics studying a trend in my dataset. My data shows quarterly unemployment and inflation for one country over 20 years, both rising together. Explain what that pattern suggests, and note two limitations of drawing conclusions from this data alone. Bullet points, under 200 words."
"You are a PhD student in political science. I'm writing a paper comparing how presidential and parliamentary systems handle legislative gridlock. Summarize the key structural differences relevant to gridlock, formatted as a short comparison table."
"You are a PhD student in public affairs writing a policy memo. My topic is the effect of state minimum-wage increases on small-business employment in a mid-sized U.S. metro. Draft the problem statement and three evidence-based options, each with one trade-off. Under 250 words, formal memo tone."
"You are a student in social data analytics analyzing social media data. My dataset is geotagged posts about a local election over one month. Suggest three text-based analyses (sentiment, topic, network) that would reveal patterns in public discourse, and note one validity concern with inferring public opinion from this kind of data. Bullet points, under 200 words."
"You are a student in international political economy. I'm writing a paper on how trade dependence affects a country's foreign policy alignment. Summarize the key theoretical mechanisms, formatted as a short comparison table of competing explanations, and note which has stronger empirical support in the literature."
Notice: every version still runs on the same five pieces — role, task, context, format, and an example where it helps.
Watch For This
Common failure modes
Vague verbs — "help me with," "look at this"
No role — model defaults to a generic assistant voice
No constraints — leads to long, unfocused answers
Assuming the model knows your dataset or field conventions
"One-shotting" — trying to nail a complex task in a single try instead of refining turn by turn
Closing the Loop
Prompting well doesn't replace verifying
- A better prompt gets you a better first draft
- It does not get you a checked one — that part is still yours
Carry This Forward
The same five-part discipline — role, task, context, format, example — applies to every prompt you write for R code today.