How to Write Better Prompts
CSE 5 – Principles of AI | Discussion Session
Blind Taste Test
Prompt Autopsy
Chain of Thought
Meta-Prompting Live
Peer Prompt Swap
Techniques covered
Role + Context
Output Format
Few-shot Examples
Chain of Thought
Negative Constraints
Meta-Prompting
What we will cover today
CSE 5 – Discussion Session
1
Role + Context
Activities 1 & 5
Telling AI who it is, who you are, and what the situation is. The single highest-leverage change you can make to any prompt.
2
Output Format
Activity 2
Specifying structure, length, and what to exclude forces the model to make editorial decisions aligned with your actual needs.
3
Few-shot Examples
Activity 5
One concrete example communicates style, tone, and structure faster than any written description, especially for writing tasks.
4
Chain of Thought
Activity 3
Asking the model to reason step by step before answering dramatically improves quality on any task involving judgment or argument.
5
Negative Constraints
Activities 1 & 4
Telling AI what NOT to do overrides its defaults toward diplomatic, balanced, hedged answers. Often more powerful than positive instructions.
6
Meta-Prompting
Activity 4
Using AI to build or improve your own prompts. Type 1: polish a draft. Type 2: let AI interview you and write the prompt for you.
AI tool survey results
Quick Check-in — Your Prompting Tricks
Activity 1 — Blind Taste Test (8 min)
CSE 5 – Discussion Session
Same question. Two prompts. Which output is more useful for writing a response paper?
Output A — basic prompt
The relationship between social media and mental health is complex. Some studies show correlations between heavy use and increased anxiety and depression, particularly among teenage girls. However, other researchers argue the effects are small or that causation hasn't been established. Social media can provide community for isolated individuals, but also exposes users to cyberbullying and unrealistic comparisons. Most experts recommend mindful use and further longitudinal research to better understand these relationships.
Output B — improved prompt
Step 1 — To establish causation, we'd need randomized controlled trials or natural experiments like sudden platform shutdowns. Correlational studies can't rule out that depressed teens seek social media more, not the reverse.
Step 2 — Two confounders: (a) Pre-existing mental health vulnerability — anxious teens may be heavier users. (b) Broader post-2012 stressors — economic precarity, climate anxiety — that coincide with both phone use and mental health decline.
Step 3 — The most rigorous work (Orben & Przybylski, 2019) found effect sizes smaller than wearing glasses. Haidt finds stronger effects for girls and passive scrolling, but critics note his studies are still correlational.
Verdict: Social media is probably a contributing factor for vulnerable adolescents — especially girls using passive consumption features — but the evidence does not support calling it a primary cause of the mental health crisis.
Vote first — then we reveal what changed in the prompt
Activity 1 — The Reveal
CSE 5 – Discussion Session
Prompt A
Does social media cause mental health problems?
Prompt B
[Same question]��Before answering, reason step by step:�1. What evidence would establish causation?
2. Name two confounding variables.
3. What does peer-reviewed research say?
Give a clear verdict. Do NOT say 'it's complicated' — take a position.
What changed — and why it matters:
1
Chain of Thought
Forced step-by-step reasoning surfaces assumptions the direct answer skips
2
Negative constraint
"Do not hedge" overrides the model's default toward diplomatic non-answers
3
Implicit role
Framing as an academic question shifts register from Wikipedia to methodology critique
Techniques: Chain of Thought + Negative Constraints
Activity 2 — Prompt Analysis (10 min)
CSE 5 – Discussion Session
Scenario: a student writing a policy-analysis essay that needs a clear argumentative position typed a prompt and got this. Their prof flagged it as 'summary, not analysis.' Diagnose what went wrong.
AI output (the bad one):
This is an interesting policy proposal with both merits and challenges. On one hand, banning cars could significantly reduce air pollution and carbon emissions. Many European cities like Oslo and Amsterdam have implemented car-free zones with positive results. On the other hand, such a ban raises concerns about accessibility for people with disabilities and businesses that rely on deliveries. It could also push traffic to surrounding neighborhoods. Success depends heavily on context, available alternatives, and how "quality of life" is defined.
Your group's task:
1
Name 3 things wrong with this output
Be specific — not just 'too vague'
2
Reconstruct the original prompt
What did someone probably type?
3
Rewrite the prompt
Rewrite so the AI produces a logical analysis that takes a position — not balanced pros/cons
Activity 2 — What a Good Prompt Gets You
CSE 5 – Discussion Session
Improved prompt: specify the analytical framework, name what you want, forbid summary
Rewritten prompt
Analyze the logical structure of this argument: "Cities should ban cars from downtown areas to reduce pollution and improve quality of life."
Identify exactly:
(1) The central claim
(2) Unstated assumptions it depends on
(3) One logical fallacy by name
(4) The strongest counterargument
Do NOT summarize. Do NOT say it's complex. Only analyze the logic.
Real Claude output:
(1) Central claim: Banning cars will produce two distinct goods — reduced pollution and improved quality of life — with the ban as the sufficient mechanism for both.
(2) Unstated assumptions: That adequate alternative transport already exists; that "quality of life" improves uniformly for all users including disabled residents; that pollution reduction follows rather than simply displacing traffic to adjacent streets.
(3) Logical weak point — false equivalence: The argument treats pollution reduction and quality-of-life improvement as if they follow from the same mechanism. Pollution may decrease downtown while increasing city-wide if drivers reroute. Quality of life for car-dependent residents may worsen even as pedestrian experience improves.
(4) Strongest counterargument: "Concentrated benefits, diffuse costs" — visible improvements in the downtown core are politically legible, while accessibility losses imposed on disabled residents and delivery-dependent businesses are dispersed and harder to quantify.
Key moves:
Numbered output format forces structure
Naming fallacy by name triggers specific knowledge
"Do NOT summarize" cuts the filler
Activity 3 — Chain of Thought Challenge (8 min)
CSE 5 – Discussion Session
Try both versions on your phone right now. Then compare with a partner.
Version A — no CoT
A historian argues that the Industrial Revolution was the primary cause of modern inequality. A sociologist argues it was colonialism. Who has the stronger argument?
Real output ↓
Both arguments have significant merit, and most contemporary scholars would argue they are deeply intertwined rather than competing. The historian highlights how mechanization created the capital-labor divide we still see today. The sociologist points to how wealth extraction from colonized regions funded industrialization and how post-colonial dependencies persist. The strongest position is likely integrative: the Industrial Revolution and colonialism were co-constitutive processes. Choosing one as "primary" may reflect disciplinary bias more than historical reality. Both scholars are probably right about their piece, but potentially wrong to claim primacy.
Version B — with CoT
[Same question]
Before giving your verdict, reason step by step:
1. What standard of evidence makes one cause "primary"?
2. What does each explanation account for well — and fail to explain?
3. Are these competing or complementary?
Then give your actual verdict. Do not hedge.
Real output ↓
Step 1 — A cause is "primary" if it (a) is logically independent of the other, (b) has greater explanatory breadth across dimensions of inequality, and (c) removing it counterfactually would collapse the other's effects.
Step 2 — Industrial Revolution: explains class stratification and capital accumulation within nations. Fails to explain racial inequality within those nations, or why the wealthiest industrial powers were the same ones with the largest colonial empires.
Step 3 — Colonialism: explains the global North-South wealth gap, the racial wealth gap within former colonial powers, and why certain inequalities map almost perfectly onto colonial history. Weaker at explaining inequality within nations that industrialized without significant colonial holdings.
Step 4 — Not fully competing: colonialism funded early industrialization (Eric Williams' thesis). But colonialism is also causally prior.
Verdict: The sociologist has the stronger argument. Colonialism has greater explanatory breadth — accounting for both global and racial inequality the Industrial Revolution framework cannot reach — and it is causally prior, having generated the capital that made British industrialization possible.
Activity 4 — Meta-Prompting Live
CSE 5 – Discussion Session
Type 1 — Prompt Polish
You have a draft. AI refines it.
Role: AI as editor
Try this prompt:
Here's my draft prompt:
"Give me feedback on my essay."
Improve it. Add specificity, useful constraints, and prevent generic responses. Keep it under 100 words.
Type 2 — Prompt Architect
You don't know what you want. AI interviews you.
Role: AI as consultant
Try this prompt:
I need a reusable prompt for a task I do regularly, but I can't phrase it well.
Ask me questions one at a time to understand what I need. Don't give answers yet. When you have enough, say "Ready" — then write the best possible prompt.
The task: critiquing arguments in a paper I'm reviewing.
Key distinction: Type 1 = you know what you want. Type 2 = AI figures it out by asking. Type 2 is a role reversal — the AI becomes the interviewer.
Bonus Demo — Role Extremity: Attack My Argument
CSE 5 – Discussion Session
Same claim. Two completely different postures. Watch what changes.
Prompt A — ask for counterarguments
"Social media algorithms are primarily responsible for political polarization in the US."
What are some counterarguments?
Real output:
Several important counterarguments exist. First, polarization in the US predates social media — partisan divides were increasing throughout the 1990s before Facebook and Twitter existed. Second, countries with similar social media usage show different polarization levels. Third, research by Boxell et al. found polarization increased most among older Americans who use social media least. Fourth, the "filter bubble" thesis has been challenged — most people do encounter cross-cutting content. Fifth, structural factors like income inequality, geographic sorting, and media economics may be more fundamental drivers.
Prompt B — hostile reviewer persona
[Same claim]
Attack this as hard as possible. You are a hostile reviewer who thinks this argument is fundamentally flawed. Find the single most devastating weakness. End with one sentence that, if true, would destroy my argument.
Real output:
This argument has a fatal empirical problem: Boxell, Gentzkow, and Shapiro showed polarization increased most rapidly among Americans over 65 — the demographic least likely to use social media. If algorithms were the primary driver, we'd expect the opposite. You cannot dismiss this as a confound; it directly falsifies the mechanism your argument depends on.
Furthermore, "primarily responsible" requires you to outweigh decades of evidence on geographic sorting, elite-driven realignment, and economic anxiety. You've asserted, not argued.
The sentence that destroys your argument: If polarization accelerated fastest among 65+ Americans who use social media at one-fifth the rate of young adults, then whatever drives polarization is largely independent of algorithms.
Technique: role + persona + negative constraint ('do not be balanced') + specific deliverable
Activity 5 — Peer Prompt Swap (7 min)
CSE 5 – Discussion Session
Start from the same base. Improve it for your own discipline.
Starting prompt for everyone:
"Explain the concept of [fill in your field's core concept] to someone who has never taken a class in your field."
1
Fill in your concept
Pick something central to your major — not too broad, not trivial
2
Pass to your partner
They can ONLY add — no deleting or replacing anything
3
Run both versions
Paste original and revised into ChatGPT/Claude right now
4
Share the one addition that mattered most
What single sentence changed the output the most?
Closing question: what did you add — and what does that tell you about what was missing?
Your Prompt Template
Works for: ChatGPT · Claude · Gemini · Midjourney · Claude Code
Role:
You are a [role] helping a [your background] student/researcher.
Context:
I am working on [task]. The audience is [who]. Stakes: [why it matters].
Format:
Respond as [structure]. Keep it under [length]. Do NOT [thing you hate].
Example:
Here's the style/tone I want: "[paste one example]"
Reasoning:
Think step by step before answering. Show your reasoning.
Question:
[Your actual question here]
You don't need all 6 every time — but each one you add narrows the gap between what AI gives and what you actually need.