Maya is a junior at Ohio State, applying for a summer marketing internship at a mid-size design agency. She's smart, funny in interviews, and her portfolio is genuinely good. But she's been rejected from twelve applications so far, and she finally asks a friend to look at her cover letter.
Her friend reads the first line: "I am writing to express my sincere interest in the Marketing Intern position at your esteemed organization." She looks up. "Did you use ChatGPT?" Maya nods. "Did you give it anything to work with?" Maya stares blankly.
The problem wasn't that she used AI. It was that she handed it a blank page and expected it to know who she was. The AI did what AI does when it has nothing specific — it wrote the most statistically average cover letter it could produce. Inoffensive. Invisible. Forgettable.
Language models are trained to be plausible in the broadest sense. When you say "write me a cover letter for a marketing internship," you haven't given it a persona, a voice, a differentiating detail, or a target. You've given it a category. And categories produce category-average output.
Think of it this way: if you asked a talented ghostwriter to write your cover letter but refused to tell them anything about you — your projects, your style, your actual reason for wanting this role — you'd get something generic too. The AI isn't broken. Your prompt is just missing the raw material.
The fix is to treat a writing prompt like a creative brief. A creative brief has a subject (who is writing), a purpose (what this piece needs to do), an audience (who reads it and what they care about), constraints (length, tone, format), and at least one specific detail that makes it personal.
Generic prompt → generic output. Specific prompt → specific, usable output. The specificity you put in directly determines the usefulness you get out. This is less a rule of AI and more a rule of communication.
Here are the five elements that turn a generic writing prompt into something that produces real output:
❌ Weak: "Write a cover letter for a marketing internship at a design agency."
vs.
✓ Strong: "Write a cover letter for a summer marketing internship at a boutique design agency called Fieldwork Studio. The reader is a creative director who is skeptical of people who haven't done real work. I'm a junior at Ohio State studying communication. My biggest relevant achievement: I ran a campus Instagram account that grew from 400 to 2,100 followers in one semester by shifting to behind-the-scenes video content. My writing voice is direct and slightly informal — here are two sentences I wrote myself: [paste samples]. Keep it under 250 words. Don't use the word 'passionate.' End by asking for a 20-minute call."
Most people send one prompt, get one output, and call it done. The people who actually get good results treat prompting as a conversation. You get a draft, you identify what's off — too formal, buries the key detail, weak opening — and you say exactly that in a follow-up prompt.
"The opening is too safe. Rewrite just the first two sentences to start with the Instagram story instead of a statement about my interest in marketing." That kind of targeted surgical feedback produces dramatically better second drafts than "make it better."
Here's what your peer group is actually doing: most people prompt once, tweak manually, submit. A smaller group iterates two or three times. An even smaller group iterates and also gives the AI a specific persona or voice reference. That last group consistently gets better outputs — and it's not a large gap in effort.
Before you write your next AI writing prompt, write a three-line brief first: who you are (including a voice sample), what the piece needs to accomplish, and one specific real detail that makes you different. That brief becomes your prompt. Your output will be unrecognizable compared to the generic version.
AI can match tone, structure arguments, vary sentence length, and produce fluent prose. What it cannot produce from nothing: genuine insight about your actual experience, authentic stakes (why you actually want this thing), and the kind of surprising specificity that comes from a real memory. These are inputs you have to supply.
If you treat AI like a writing partner who needs to be briefed — rather than a vending machine that dispenses essays — the quality of what you get back changes fundamentally. Your job shifts from "person who types a request" to "person who knows what good looks like and directs toward it." That's actually a more interesting role.
You're going to build a writing brief for a real piece you need — cover letter, email to a professor, cold message to someone you want to connect with, personal statement, anything. Share your brief with the AI advisor and get honest feedback on whether it has everything needed to produce output that's actually yours.
The advisor will push back if your brief is too vague, ask you for the details you skipped, and tell you when something is actually good. Don't expect cheerleading.
Darius is a second-year CS student at Georgia Tech. He's building a Flask web app for a class project — a simple grade tracker. He hits a bug: his database isn't saving correctly. He pastes the broken function into Claude and says "fix this."
Claude returns a revised function. Darius copies it in. The original bug is gone. But now there's a different error he's never seen before — something about a SQLAlchemy session scope. He pastes that into Claude. It returns more code. He pastes that in. Three hours later, he has a working app and zero understanding of what any of the last fifteen functions he added actually do.
Then his TA asks him to walk through his code in office hours. It doesn't go well.
Darius's problem isn't that he used AI. It's that he used it in a way that outsourced not just the typing but the thinking. Every time you paste code in without understanding what it does, you're borrowing a solution without understanding the debt — and in coding, that debt compounds fast.
The alternative isn't to refuse AI help. It's to prompt in a way that keeps you in the loop. This means asking for explanations alongside code, asking why a particular approach was chosen, and asking what you'd need to understand to maintain or modify this code later.
The key distinction: AI as pair programmer vs. AI as outsourced dev. In the first mode, you're the one who understands what's being built and why — the AI is producing implementations that you verify and learn from. In the second mode, you've handed off the thinking entirely. The first mode makes you better. The second makes you dependent.
A significant portion of your peers are in full copy-paste mode. They pass the assignment, fail the interview, and can't debug anything they didn't write from scratch. This isn't a moral judgment — it's a practical one. The skill gap shows up fast in environments where the AI isn't available or where you have to modify the code six weeks later.
AI-generated code fails most often when it doesn't understand your environment. The model doesn't know you're on Python 3.11, using Flask 3.0, with SQLAlchemy in a specific configuration, targeting a SQLite database, with a particular project structure. When you omit context, it guesses — and its guess might be plausible but wrong for your setup.
❌ Weak: "This code is broken, fix it." [pastes 20 lines with no context]
vs.
✓ Strong: "I'm using Python 3.11 with Flask 3.0 and SQLAlchemy 2.0. This function is supposed to save a new grade entry to a SQLite database. When I call it, I get: [paste exact error message]. Here's the function and the model it writes to: [paste code]. Don't add new dependencies. After you fix it, explain in one paragraph why the bug occurred and what your fix does differently."
These are different use cases that need different prompt structures.
Debugging prompts work best when you include: the error message (exact text), the code that caused it, what you expected to happen, and what actually happened. The more precisely you describe the gap between expected and actual behavior, the more targeted the fix.
Build prompts — where you're asking AI to generate something from scratch — need a clear specification: what inputs the function or component receives, what it returns or renders, any edge cases it needs to handle, and what it should explicitly not do. Think of this as writing a spec before asking for an implementation.
After AI gives you code, add one line to your prompt before you accept it: "Now explain this code in two sentences as if I'm going to have to modify it in three weeks." If you can't follow the explanation, that's a signal you need to dig deeper before moving on — not that you should just move on.
There are situations where AI-assisted code is actively counterproductive: when you're learning a new concept for the first time and the struggle is the point; when you're in an interview setting and the expectation is that you understand; when you're working in a security-sensitive codebase where you can't paste code to external services; and when the code is so context-specific that the model's plausible-but-wrong suggestions will cost you more time than writing it yourself.
The honest version of "AI makes me a faster developer" is: AI makes you faster at things you already understand well enough to verify. For things you don't understand, AI can make you faster at producing broken code you can't diagnose. That's not faster — it's just a different kind of slow.
Bring a real coding problem — something you're actually working on, or a realistic scenario. Describe it to the advisor and get help crafting a prompt that will produce code you can actually understand and maintain. The advisor will ask about your environment, your constraints, and what you already understand about the problem before helping you build the prompt.
The goal isn't just a working answer — it's a prompt strategy you can replicate. The advisor may also ask you to explain a piece of code to verify comprehension.
Jordan is writing a policy brief for a political science seminar on the effects of rent control on housing supply. Deadline is in 36 hours. She asks ChatGPT: "What does the research say about rent control and housing supply?"
ChatGPT gives her a fluid, confident four-paragraph summary citing multiple economists and a few studies. She writes her brief around it. Her professor hands it back with a note: two of the studies don't appear to exist, one economist is misattributed, and the paper she described as "widely cited" doesn't show up in any database he checked.
Jordan is furious at the AI. But the AI never claimed to be a library. She asked it to summarize research; it synthesized plausible-sounding research from its training patterns. She didn't ask it to distinguish what it actually knew from what it was generating. That's the prompt gap.
When you ask an AI to summarize research on a topic, you're not querying a database of papers. You're asking the model to generate text that sounds like an accurate research summary. Those are radically different things. The model has no index of verified papers to pull from — it has patterns of how academic summaries sound, and it fills those patterns with plausible-sounding content.
This means AI is excellent for some research tasks and genuinely dangerous for others. The key is knowing which is which.
Explaining concepts and mechanisms you can then verify. Giving you the vocabulary and frameworks to search the actual literature. Helping you identify what questions to ask. Summarizing documents you paste to it. Stress-testing your argument by generating counterarguments.
Citing specific studies with accurate titles, authors, and findings. Knowing what's happened in a field after its training cutoff. Distinguishing between well-established consensus and contested fringe claims. Knowing when it's confabulating vs. recalling.
The fix is to change what you're asking for. Instead of asking AI to tell you what the research says, ask it to help you understand the landscape of the debate, then do your own verification in actual databases (Google Scholar, JSTOR, PubMed, Semantic Scholar).
There's a specific prompt technique that dramatically reduces hallucination risk: ask the AI to flag its own uncertainty.
"When answering this question about rent control research, explicitly flag any claims where you're less than confident in the accuracy. Use phrases like 'I believe but am not certain' or 'you should verify this.' I'd rather have an honest uncertain answer than a confident wrong one."
This doesn't eliminate hallucination, but it changes the model's behavior in a useful direction. Most models respond to this kind of explicit instruction by hedging more where they actually are uncertain. It's not a guarantee — but it's better than asking for confident summaries and getting confident fiction.
Most of your peers who use AI for research are treating it like a search engine with better sentences. The ones who get real utility out of it treat it like a smart study partner who is great at explaining concepts but who you don't trust to cite papers from memory. That's a useful peer-accurate framing.
Never cite a source you found through AI without independently verifying it exists in an actual database. Use AI to understand concepts and the structure of debates, then go find real sources using the vocabulary the AI gave you. That combination — AI for conceptual orientation, databases for citations — is faster and more accurate than either alone.
Research prompts also apply to non-academic tasks: researching a company before an interview, understanding a financial product, learning about a neighborhood you might move to. The same principles apply, but the hallucination risk changes by domain.
For career research, AI is excellent at explaining how an industry works, what roles involve, what skills are valued, and what questions to ask in an interview. It's much weaker on specific company details, recent news, and anything that requires knowing what's happened in the last year or two. For that, you need the company's actual site, recent news, and LinkedIn.
The habit to build: treat AI as your orientation layer, not your verification layer. It gets you oriented in a new domain quickly. Verification happens through primary sources.
Bring a research question you're actually working on — a paper topic, an industry you're trying to understand, a policy question you're exploring, a company you're researching. The advisor will help you build a research strategy that uses AI for what it's actually good at while directing you to primary sources for what it isn't.
The advisor will also play adversarial auditor — if you share an AI-generated claim, it will push you on whether you've verified it and where you'd go to do that.
Priya is trying to decide between two job offers. One pays more but is at a company she doesn't know much about. One pays less but is at a well-known company with a good brand name. She asks ChatGPT: "I got two job offers — one pays more, one has a better brand. Which should I take?"
ChatGPT gives her a balanced response. It lists pros and cons of each. It ends with: "Ultimately, it depends on your priorities and values. Both are valid choices!" Priya feels slightly more confused than before.
The problem is that she asked AI for a decision when she should have asked it to help her think through one. She also gave it almost no information about her situation. The AI responded to the prompt it was given — a content-free decision question — with content-free balanced analysis. Garbage in, diplomatic mush out.
People use AI for decision support in three ineffective ways: they ask for a recommendation without giving enough context, they ask for pros and cons (which they already know), and they ask leading questions where the expected answer is already implied. All three produce useless or actively harmful outputs.
AI is genuinely useful for decisions in three different ways that most people don't use:
AI has a tendency toward agreement and balance — it's trained to be helpful and non-confrontational. When you ask questions that signal what answer you want, it often delivers that answer. The fix is to explicitly ask for pushback: "I expect you to disagree with me here, not validate. Tell me where my thinking is wrong."
Analysis prompts — for business decisions, policy thinking, financial choices, creative strategy — work best when you treat AI as a structured thinking partner rather than a summary generator.
❌ Weak: "What are the pros and cons of starting a clothing brand on Instagram vs. TikTok?"
vs.
✓ Strong: "I'm considering launching a vintage clothing resale brand. My target customer is women 18–28 who already buy from Depop and Vinted. I have $500 to spend on content in the first 90 days. I'm leaning toward TikTok because that's where I see engagement happening in this space. What's the strongest argument against TikTok and for Instagram instead? Assume I'm motivated enough that enthusiasm isn't the variable — the real question is platform-market fit and my specific resource constraints."
The difference: the strong prompt gives the AI your specific context, your tentative conclusion, and a pointed question that asks for the counterargument. It also pre-empts the generic "both are valid!" response by specifying what the actual variable is.
Creative prompts have a different set of failure modes. The most common: asking for too much at once ("write a short film script"), accepting the first draft without iteration, and not giving the AI the aesthetic reference points it needs.
Creative AI use works best as a collaboration, not a delegation. The model generates; you curate, direct, and refine. This mirrors how good creative direction works — a director doesn't do everything, but they have a clear vision of what "right" looks like.
For any decision or analysis, the most useful prompt you can write asks AI to disagree with you, not confirm you. Explicitly say: "I'm leaning toward X — give me the strongest case against it." For creative work: give references, not adjectives; give constraints, not blank pages; iterate with specific direction, not vague requests for improvement.
The underlying skill across all four domains in this module — writing, coding, research, and analysis/creative — is the same: knowing what kind of output you need before you ask. Are you looking for a draft to react to? A challenge to your existing view? A structured format to fill in? Vocabulary to search with? An implementation to verify? Each of these needs a different prompt architecture.
Most people who feel like AI "isn't that useful" are in the habit of asking AI what it can do for them, then reacting to the output. The people who get real leverage from it know what they need going in — and prompt specifically for that thing. That's a thinking discipline, not a prompting technique. And it transfers everywhere.
Bring a real decision you're working through — a career choice, a financial decision, a creative project direction, a business idea — or a creative brief you want to develop. The advisor will not validate your existing thinking. It will ask what you're assuming, make the case for alternatives you're not considering, and challenge the weakest parts of your reasoning.
If you bring a creative brief, the advisor will push you to give references instead of adjectives and constraints instead of open-ended requests. Expect pushback. That's the point.