How to Write a Prompt That Actually Works: One Bad Prompt, Five Rewrites

This guide is for anyone who has typed a question into ChatGPT or Claude, read the answer, and thought that's it? — students, developers, writers, marketers, founders, and anyone who suspects the tool is smarter than the results they're getting. We're going to take one genuinely bad prompt and rewrite it five times. You'll see the actual output at every stage, and by the end you'll have a formula you can reuse, a library of copy-paste prompts, and a much better instinct for why AI gives you boring answers. Reading time is about fourteen minutes. The part where it clicks usually happens around rewrite three.

Here's the thing that changed how I use these tools. For most of my first year with AI, I thought I was evaluating the model. I'd type something in, get something mediocre back, and quietly file it under "overhyped." It took me embarrassingly long to notice that the model was doing exactly what I asked. I just hadn't asked for much.

The Cover Letter That Broke My Confidence

A friend asked me to help with a job application. Easy, I thought. I've been using this stuff for months.

I typed: Write a cover letter.

What came back was — you already know what came back. "I am writing to express my strong interest in the position." Three paragraphs about being a passionate, results-driven professional. A closing line about how they'd welcome the opportunity to discuss further. It was grammatically flawless and completely unusable. It could have been sent by anyone, to anyone, for any job.

My first reaction was to blame the tool. My second, about ten minutes later, was worse: I realised that if a human freelancer had received my request — three words, no job description, no name, no company, no idea what my friend had ever done for a living — they would have produced exactly the same thing. Or more likely, they'd have replied with six questions. The AI can't ask the six questions. So it guesses, and it guesses average, because average is the safest guess when you've told it nothing.

That's the whole article, really. But knowing it and feeling it are different, so let's run the experiment.

I'm going to take one lazy prompt and improve it five times. Same tool, same session, one change per rewrite. You'll see every output. Some of them are still bad — I've kept them honest rather than making the progression look neater than it was.

What's In This Guide

• Why most AI prompts fail
• The one bad prompt (and its output)
• Rewrite 1 — Add context
• Rewrite 2 — Define the audience
• Rewrite 3 — Add structure
• Rewrite 4 — Add tone and constraints
• Rewrite 5 — Give it a role
• Side-by-side comparison
• Why this happens
• The universal prompt formula
• Copy-paste prompts for real jobs
• Common mistakes
• Advanced tricks worth learning
• Case studies
• My own worst prompting habits
• FAQ

Why Most AI Prompts Fail

Almost every bad prompt I've seen — mine included — is missing at least one of the same seven things.

It's vague. "Write an essay." On what? For whom? How long? Every unanswered question becomes a coin flip the model makes without you.

It has no audience. Writing for a professor, a recruiter, and a fourteen-year-old are three different jobs. Unspecified means "generic adult," which is nobody.

It has no objective. "Make a resume" doesn't say whether the goal is to get past a screening filter, to impress a hiring manager, or to fit on one page. Those pull in different directions.

It has no context. You know your situation. The model doesn't and can't. "Explain Python" gives you the same output whether you're a complete beginner or a Java developer switching languages, because you didn't say which.

It has no constraints. No length, no things to avoid, no boundaries. So you get whatever length feels natural and whatever clichés are statistically popular.

It has no examples. If you know what good looks like and don't show it, you're asking the model to reverse-engineer your taste from nothing.

It has no format. Bullets or paragraphs? Table or prose? Unspecified means you'll spend five minutes reformatting.

Look at the prompts most people actually type: Write an essay. Make a resume. Explain Python. Create marketing. Every one of them fails on at least five counts. And then we conclude the AI isn't that good.

The One Bad Prompt

Here's our starting point. I picked it because it's the kind of thing I'd genuinely type at 11 p.m. when I want something done and don't want to think.

Prompt v0: Write a LinkedIn post about AI.

Six words. Grammatically fine. Practically useless, and it's worth naming exactly why: there are roughly infinite LinkedIn posts about AI that would satisfy this request. The model has to pick one, and it has no information to pick well, so it picks the middle of the distribution.

Output #1

🚀 AI is transforming the way we work!

In today's fast-paced digital landscape, artificial intelligence is no longer a futuristic concept — it's here, and it's reshaping industries across the globe.

From automating repetitive tasks to unlocking powerful insights from data, AI is empowering businesses and professionals to achieve more than ever before.

But here's the real question: are we ready to embrace this change?

The companies that thrive tomorrow will be the ones that adapt today. 💡

What are your thoughts on AI in the workplace? Let me know in the comments! 👇

#AI #ArtificialIntelligence #Innovation #FutureOfWork #Technology

You've read this post four hundred times. It has a rocket emoji, a rhetorical question, a line about the fast-paced digital landscape, and a call to action that nobody has ever acted on.

So why did it do this? Not because it's dumb. Because I asked for the average LinkedIn post about AI, and this is the average LinkedIn post about AI. The model has absorbed millions of them. Given no reason to deviate, it produced the centre of that pile. It answered my question correctly. My question was the problem.

Rewrite #1 — Add Context

The first fix is the biggest and the one people skip most: tell it what actually happened.

Prompt v1: Write a LinkedIn post about AI. Context: last month my team replaced our manual invoice-checking process with an AI tool. It reads scanned invoices and flags mismatches against our purchase orders. It cut a two-day task to about twenty minutes, but it also missed three duplicate invoices in the first week and we had to add a human review step. I want to post about what we learned.

Output #2

What we learned automating invoice checks with AI

Last month, our team replaced a manual invoice-checking process with an AI tool that reads scanned invoices and flags mismatches against our purchase orders.

The results were impressive: a task that used to take two days now takes around 20 minutes.

But it wasn't perfect. In the first week, the tool missed three duplicate invoices. We responded by adding a human review step to catch what the AI couldn't.

Our key takeaways:
✅ AI can dramatically reduce time spent on repetitive tasks
✅ Automation is not the same as accuracy
✅ Human oversight remains essential

AI is a powerful tool, but it works best alongside people, not instead of them.

What has your experience with AI automation been? 👇

#AI #Automation #Finance #DigitalTransformation

What changed: it now has something to say. There are real details in there — two days to twenty minutes, three missed duplicates, the review step. Nobody else could have written this post, which is the first time that's been true.

What didn't change: the shape. Still tick-box emojis, still a "key takeaways" list, still the participation-trophy question at the end, still "AI is a powerful tool but it works best alongside people," which is the single most recycled sentence in the genre. The facts are mine. The voice is still the internet's average.

Context bought us substance. It didn't buy us a voice. That's the next problem.

Rewrite #2 — Define the Audience

Here's a question I never used to ask myself before writing anything: who is supposed to read this, and what do they already know?

Right now the model is writing for "LinkedIn" — which means everyone, which means nobody. Let's fix that.

Prompt v2: [same context as before]

Audience: finance managers at mid-sized companies who are being pushed by their leadership to "use AI" and are quietly nervous about it. They understand invoices and approval workflows deeply. They do not care about AI as a technology trend and they've heard the transformation talk already. What they want to know is whether this is worth their time and what it will cost them in problems.

Output #3

If your CFO is asking about AI, here's what actually happened when we tried it.

We automated invoice checking last month. Here's the honest version.

The tool reads scanned invoices and flags mismatches against our POs. What took two days now takes about 20 minutes. That part is real and I'd do it again.

But in week one it missed three duplicate invoices. Not mismatched — duplicated. The kind of thing that quietly becomes a payment run problem if nobody catches it.

So we added a review step. One person, 15 minutes a day, checking the flags and spot-checking what wasn't flagged. That's the part nobody puts in the case study.

Net position: still a large win. Two days to about 35 minutes including review. But "we automated it" is not the same as "we stopped looking at it," and if you're being asked to roll something like this out, budget for the review step before someone asks you why a duplicate went through.

Happy to share what we checked before buying if that's useful.

#Finance #AccountsPayable #AI

Now we're getting somewhere. Notice what disappeared without me asking: the rocket emoji, the tick boxes, the "in today's landscape" opener. Nothing in my prompt said "remove the emojis." Defining a skeptical, busy, domain-expert audience did it, because that audience wouldn't tolerate them and the model knows that.

Notice also what appeared: "the kind of thing that quietly becomes a payment run problem." That's a sentence written for someone who knows what a payment run is. And the closing line changed from a generic engagement bait question into a specific offer, which is a completely different move.

Audience is the highest-leverage line in a prompt after context, and almost nobody includes it.

Rewrite #3 — Add Structure

The output above is good but shapeless. It's a wall of short paragraphs, and on a phone that's a lot of scrolling before you hit the point. LinkedIn also cuts posts off after roughly two lines with a "see more" link, so the opening carries a disproportionate load.

Let's ask for a shape.

Prompt v3: [same context and audience]

Structure:
- Line 1: a specific, concrete hook that works standalone when LinkedIn truncates the post. No questions, no statistics.
- Then: what we did, in two lines.
- Then: the failure, in detail. This is the most important section — give it the most room.
- Then: how we fixed it and what it cost.
- Final line: one clear takeaway a reader can act on tomorrow.
- Short paragraphs, most 1–2 lines. Blank line between each.

Output #4

Our AI invoice checker missed three duplicate invoices in its first week.

We still kept it. Here's why.

Last month we replaced manual invoice checking with a tool that reads scans and flags mismatches against our POs.

Two days of work became twenty minutes. That number is real.

Then week one happened.

Three duplicates went through clean. Not flagged, not queried — just passed, because the tool was trained to catch mismatches and a perfect duplicate isn't a mismatch. It's two identical correct invoices. The system had no reason to blink.

That's the failure mode nobody warns you about: it isn't wrong, it's narrow. It does the job you specified, not the job you meant.

Our fix was unglamorous. One person, fifteen minutes each morning, reviewing flags and spot-checking a sample of what wasn't flagged. Duplicates now get caught there.

Real number: two days → about thirty-five minutes, review included.

If you're rolling out something similar, write down what the tool is not looking for before you switch off the manual process. That list is your review step.

What structure bought us. That opening line is doing serious work — it's specific, slightly alarming, and survives truncation. Compare it to "AI is transforming the way we work!" from output one and the distance is almost funny.

But look at the middle section, because that's the real gain. By telling the model the failure deserved the most room, it actually thought about the failure — and produced the best insight in the entire post: it isn't wrong, it's narrow. A perfect duplicate isn't a mismatch. I hadn't articulated that myself. The structure forced depth where I wanted depth, and depth is where insight lives.

Readability improved too, and that's not a cosmetic point. Short paragraphs with breathing room are how people actually read on a phone, and a post nobody finishes has no takeaway at all.

Rewrite #4 — Add Tone and Constraints

Version three still has a few tells. "That's the failure mode nobody warns you about" is a touch performative. And the post is drifting long.

Constraints are where I used to be too polite. Asking for "a human tone" does nothing — the model has no idea what you find human. You have to name the specific things you don't want.

Prompt v4: [same context, audience, structure]

Tone and constraints:
- Maximum 180 words. Hard limit.
- Write like you're telling a colleague over coffee, not presenting to a board.
- Never use: leverage, unlock, game-changing, transformative, journey, in today's landscape, revolutionize, seamless, empower.
- No emojis. No hashtags. No rhetorical questions.
- Do not end on an inspirational summary line.
- Do not claim AI works "best alongside humans" or any variation. That sentence is exhausted.
- One call to action only, and make it a specific offer rather than "share your thoughts."
- Include at least one detail that could only come from having actually done this.

Output #5

Our AI invoice checker missed three duplicate invoices in its first week. We kept it anyway.

The tool reads scanned invoices and flags mismatches against our POs. Two days of checking became twenty minutes.

Then three duplicates sailed through. Not flagged, not queried. And when I dug into it the reason was almost reasonable: the tool looks for mismatches, and two identical correct invoices don't mismatch anything. It wasn't wrong. It was narrow.

Our fix cost fifteen minutes a morning — one person reviewing flags and spot-checking a sample of what didn't get flagged. Deeply unexciting. Catches the duplicates.

Honest maths: two days → about thirty-five minutes with review included. Still worth it.

The thing I'd tell my past self: before you switch off the manual process, write down what the tool isn't looking for. We did that after the incident. It should have been the first thing on the whiteboard.

If you're evaluating one of these, I'll send you the checklist we used. Just message me.

Word count: 178. Tighter, and tighter turns out to mean better, because padding is exactly where the clichés were hiding. Cutting the word budget cut the filler automatically.

The banned-words list did more than remove words. It removed the register those words belong to. And "It wasn't wrong. It was narrow." is now a two-sentence punch instead of a sentence explaining that it's a punch.

My favourite addition is one I didn't ask for directly: "We did that after the incident. It should have been the first thing on the whiteboard." That came from the instruction requiring a detail only someone who'd done it would have. Admitting the checklist came late is a small, slightly embarrassing truth — and small embarrassing truths are what make writing sound like a person.

Rewrite #5 — Give It a Role

Last move, and this one gets over-hyped, so let me be precise about what it does. Telling a model it's an expert doesn't make it one. What a role does is narrow the vocabulary, priorities, and instincts it draws from. "You are a world-class genius" is worthless. "You are someone who has written this specific kind of thing for this specific kind of reader" is not.

Prompt v5:

ROLE: You are a finance operations lead with twelve years in accounts payable who started writing on LinkedIn two years ago. You're mildly allergic to corporate language. Your posts do well because you write about things that went wrong, in plain terms, without turning them into life lessons.

GOAL: Write one LinkedIn post that makes finance managers trust me enough to ask for the evaluation checklist.

AUDIENCE: Finance managers at mid-sized companies, being pushed by leadership to "use AI," quietly nervous. Deep expertise in invoices and approvals. Zero interest in AI as a trend. They want to know: is this worth it, and what will it cost me in problems?

CONTEXT: Last month we replaced manual invoice checking with a tool that reads scans and flags PO mismatches. Two days → 20 minutes. Week one: it missed three duplicate invoices, because a perfect duplicate is not a mismatch. We added a daily 15-minute human review. Net: two days → ~35 minutes.

STRUCTURE: Concrete hook that survives truncation → what we did (2 lines) → the failure, in the most detail → the fix and its real cost → one actionable takeaway.

CONSTRAINTS: Max 180 words. No emojis, hashtags, or rhetorical questions. Banned: leverage, unlock, game-changing, transformative, journey, seamless, empower, in today's landscape. No inspirational closing line. Never say AI "works best alongside humans." One CTA: offer the checklist. Include a detail only someone who did this would know.

QUALITY CHECK: Before answering, verify every constraint is met. If any sentence could appear in a vendor's marketing page, rewrite it.

Output #6 — Final

Three duplicate invoices went straight through our new AI checker in week one. We kept it running.

The tool reads scans and flags anything that doesn't match the PO. Two days of manual checking dropped to about twenty minutes.

The duplicates were the interesting part. Nobody flagged them because nothing was wrong with them — two identical, correct, matching invoices. The tool was looking for mismatches. A duplicate isn't a mismatch. It did precisely what we asked and precisely not what we wanted.

The fix is boring. Fifteen minutes each morning, one person, reviewing the flags and sampling what wasn't flagged. It catches them.

Real number, with review included: two days to about thirty-five minutes. I'd sign for that again.

The bit I'd do differently: we wrote down what the tool doesn't check only after it bit us. That list should exist before you turn the manual process off. It is, functionally, your review step.

I've got the checklist we used to evaluate it. Message me and I'll send it over.

Is it dramatically better than version four? Honestly — no. It's maybe fifteen percent better, and I want to say that plainly because most articles like this one pretend the final version arrives like a thunderclap. It doesn't. The big jumps were context and audience. Everything after that is refinement.

But the refinements are real. "It did precisely what we asked and precisely not what we wanted" is the sharpest line in any of the six versions, and it came from the role — a person who has been burned by systems before writes that sentence; a general-purpose assistant doesn't. The rhythm changed too. Short sentence, longer sentence, short sentence. That's a human cadence, and it showed up because I described a human who writes that way.

Would I publish this? Yes — after changing maybe two words. Version zero I wouldn't have published under a fake name.

Side-by-Side Comparison

All six versions, scored honestly. The ratings are my judgement, not a measurement.

The curve is steep early and flat late. Context and audience did most of the work. If you only ever remember two lines from this article, remember those two.

Why This Happens

A rough but useful mental model. Language models predict what text plausibly comes next, given everything before it. They're not looking up an answer — they're navigating a very large space of possible continuations and picking a likely one.

So picture your prompt as a filter:

"Write a LinkedIn post about AI"
→ millions of plausible answers
→ picks something near the middle
→ you get the average

+ context → thousands of plausible answers
+ audience → hundreds
+ structure → dozens
+ constraints → a handful
+ role → the handful, in a consistent voice
→ you get something close to what you meant

Every detail you add deletes an enormous set of wrong answers. That's the entire mechanism. There are no magic words — there's only ambiguity, and how much of it you removed before you pressed enter.

It also explains the specific failures. Constraints improve precision because "180 words, no emojis" is checkable while "make it engaging" isn't. Roles change style because they select a slice of the training data with a consistent voice. Examples work better than descriptions because a pattern is unambiguous and an adjective is a Rorschach test. Both OpenAI and Anthropic publish their own prompting guides, and they're worth reading — they say the same thing in more formal language.

The Universal Prompt Formula

Here's the skeleton I use. Not every job needs every field — for a quick question, two lines is plenty. But when output disappoints, the missing piece is almost always one of these.

ROLE — who is doing this work, described by behaviour not status
GOAL — what success looks like in one sentence
AUDIENCE — who reads it, what they know, what they don't care about
CONTEXT — the facts only you have
TASK — the specific thing to produce
CONSTRAINTS — length, banned words, hard rules
OUTPUT FORMAT — the shape you want back
STYLE — the register, described concretely
QUALITY CHECK — how it should verify its own work before answering

And a reusable template:

ROLE: You are [specific person with specific experience].
GOAL: [What this needs to achieve.]
AUDIENCE: [Who. What they already know. What they don't care about.]
CONTEXT: [Your actual situation and facts.]
TASK: [What to produce.]
CONSTRAINTS: [Length. Banned words. Hard rules.]
FORMAT: [Structure and shape.]
STYLE: [Tone, described by example not adjective.]
CHECK: Before answering, verify every constraint. If any part fails, fix it before responding.

That last line matters more than it looks. Asking the model to check its own work against the constraints you listed catches a surprising number of slips, and it costs you one sentence.

Copy-Paste Prompts for Real Jobs

All of these are stripped-down applications of the same formula. Fill in the brackets.

Students — turning a chapter into notes:

You're a tutor for [subject] preparing a student for an exam that rewards precise definitions and worked examples. Turn the material below into revision notes. Use headings per concept, a one-line definition, then a plain-English explanation, then one example. Flag anything the source states without explaining — I need to know what to ask my professor. Do not add facts that aren't in the source. Material: ---[paste]---

Students — checking your understanding:

I'm going to explain [concept] in my own words. Don't be encouraging — find every gap, imprecision, or place I've used the right word for the wrong reason. Then ask me two questions that would expose whether I actually understand it. My explanation: ---[paste]---

Developers — debugging:

Language: [X, version]. Environment: [Y]. Here's the code, the exact error, and what I've already tried. Tell me the most likely cause first with your reasoning, then two less likely causes. Don't rewrite the whole file — show me the minimal change. If you need information I haven't given, ask before guessing. ---[code]--- ---[error]--- ---[tried]---

Developers — code review:

Review this [language] code as a senior engineer who'll maintain it after I leave. Focus on: correctness bugs, edge cases I've missed, and anything a reader would misunderstand. Ignore style preferences. For each issue give the line, the problem, and the concrete failure it causes. Don't compliment the code. ---[paste]---

Resume — a specific application:

You're a hiring manager who screens sixty resumes a week for [role] and spends about twenty seconds on each. Here's my resume and the job description. Tell me: which lines you'd skim past and why, which claims you'd doubt, and the three specific edits with the biggest effect. Don't rewrite it — I want to understand the reasoning. ---[resume]--- ---[JD]---

Email — the difficult kind:

Write an email to [recipient and their position]. Situation: [what happened]. Goal: [what I need]. Constraints: under 120 words, no apology padding, no "I hope this email finds you well," one clear ask, professional but not stiff. They're busy and slightly annoyed already.

Summaries you can trust:

Summarise the text below for [audience] in under [N] words. Rules: only what's in the source, no added interpretation. After the summary, list anything important you had to leave out, and anything the source asserts without evidence. ---[paste]---

YouTube script:

You're a creator whose videos retain viewers because you make the point in the first fifteen seconds and never pad. Write a [N]-minute script on [topic] for [audience]. Structure: hook that states the payoff, the thing itself, one concrete example, a close with no subscribe-beg. No "hey guys," no "in this video we'll discuss," no throat-clearing.

Research starting point:

I'm researching [topic] and know [current level]. Give me: the five things I need to understand first and why, the two most common misconceptions, and the questions experts still disagree on. Mark anything you're uncertain about. Don't invent citations — if you're not sure a source exists, say so.

SEO outline:

Target keyword: [X]. Search intent: [Y]. Audience: [Z]. Give me an H2/H3 outline for an article that answers the intent better than the current top results. For each heading, one line on what it must cover and why a reader needs it. No filler sections. If a section only exists to hit a word count, leave it out.

Interview prep:

You're interviewing me for [role] and you're skeptical. Here's my background: [paste]. Ask me the five questions most likely to expose a weakness in it. After each answer I give, tell me what a strong candidate would have said instead. One question at a time.

Marketing copy:

Product: [X]. Audience: [Y]. The one thing they need to believe to act: [Z]. Write [format] under [N] words. Banned: revolutionary, seamless, unlock, game-changing, empower, effortless. Every claim must be specific enough that a skeptical reader could check it. One CTA.

Notice the pattern across all of them — a specific role, a real audience, a hard constraint, and something explicitly banned. That's it. That's the whole skill.

If you want more on the study-specific side of this, I've written separately about building a free AI study stack. For the resume prompts, my AI resume post goes considerably deeper.

Common Prompt Mistakes

Stacking questions. Asking five things at once gets you five shallow answers. The model spreads its effort. Ask one thing, then the next.

Being too broad. "Help me with marketing" isn't a task, it's a topic. Narrow it until it has a deliverable.

Being too short. Six words means six coin flips.

Contradicting yourself. "Be comprehensive but keep it under 100 words." It'll pick one and you'll be annoyed at whichever it picked. Decide first.

Skipping the format. Then reformatting by hand for five minutes. One line of prompt would have saved it.

No audience. The single most common omission, and the most expensive.

No constraints. Without a length limit you get padding, and padding is where the clichés live.

Describing instead of showing. If you have an example of what good looks like, paste it. Three examples beat three paragraphs of description every time.

Giving up after one try. The first output is a draft. Treat it like one.

Advanced Tricks Worth Learning

Few-shot prompting. Show two or three input-and-output examples before the real task. The most reliable way to transmit a style you can't describe. Vary your examples slightly or the model copies them too literally.

Asking for reasoning. For anything with steps — logic, maths, planning, diagnosis — asking the model to work through its reasoning before giving the answer improves accuracy, and it lets you see where it went wrong instead of just receiving a wrong answer confidently. "Work through this step by step, then give your conclusion" is the whole technique.

Iterative prompting. Don't restart when output is close. Say what's wrong. "Third paragraph is vague — replace it with a specific example. Keep everything else." Faster and more precise than rewriting the prompt.

Persona prompting. Covered above. Describe behaviour, not status.

Step prompting. For big jobs, split into stages. Outline first, approve it, then draft section by section. One-shot requests for large deliverables produce shallow everything.

Delimiters. Wrap pasted material in obvious markers: --- or <source>. Stops your content being read as instructions. Essential once prompts get long.

Checklists and rubrics. Give it your standards. "Rate this draft against these five criteria out of ten, then fix anything under eight." You get a critic and an editor for one prompt.

Self-review. Add: "Before answering, check your response against every constraint above. If any fails, fix it first." One sentence, meaningful improvement.

Real-Life Case Studies

The student who stopped generating notes. A friend in her fifth semester was producing beautiful AI notes and failing anyway. Her prompt: "Summarise this chapter." She switched to: "Turn this into ten questions an examiner would ask, with the answer hidden below each." Same chapter, same tool. The difference was that the new output made her retrieve instead of read. Notes feel like studying. Questions actually are.

The developer who kept getting the wrong fix. He'd describe bugs in his own words — "it's not saving properly" — and get plausible, useless suggestions. He started pasting the exact stack trace, the code, and what he'd already ruled out. The hit rate went from maybe one in four to most of the time. He hadn't got better at prompting in any abstract sense. He'd just stopped paraphrasing evidence.

The writer who found her voice by banning things. A blogger couldn't get drafts that sounded like her. Adjectives didn't help — "conversational," "warm," "authentic" produced the same polished mush. What worked was a list of eleven banned phrases from her own pet peeves, plus three paragraphs of her old writing as examples. Negative constraints and real samples beat positive adjectives, consistently.

The designer who asked for bad ideas. Stuck on concepts, getting the same safe suggestions. He tried: "Give me fifteen ideas. The first five should be obvious, the next five should be strange, the last five should be ideas a client would reject and explain why they'd reject them." Ideas eleven through fifteen were where the interesting work was. The permission to be wrong is what unlocked it.

The founder whose emails got replies. His cold emails were three paragraphs about his company. He rewrote the prompt around the recipient: their role, their likely week, why they'd ignore this. The output came back at four lines, one specific observation about their business, one ask. Replies went from near-zero to occasionally yes. Audience again.

The recruiter who learned what the tool wasn't looking at. She used AI to sift resumes and noticed it kept surfacing keyword-dense, experience-light candidates. Her fix was a rubric: what actually matters, what doesn't, and what to flag as suspicious. It didn't make the tool right. It made its blind spots visible — which, as our invoice example showed, is the whole game.

My Own Worst Prompting Habits

In the spirit of the article, here's what I got wrong.

I collected templates instead of learning the thinking. I had a folder of "100 mega prompts" I'd downloaded from somewhere. I used maybe four. They were fine, but they weren't mine, and a template for someone else's job is a template for the wrong job. The formula transfers. The template doesn't.

I confused politeness with clarity. I'd write "could you maybe try to make it a bit more engaging?" as if I'd hurt its feelings. The word "engaging" contains no information. Now I write "cut every sentence that doesn't contain a fact," which is rude and works.

I blamed the model for months. Genuinely. I told people it was overrated. Then I watched someone else use the same tool and get better output, and the difference was that they told it things. That was a humbling afternoon.

I trusted fluent output. This one cost me. I used an AI-written paragraph containing a confidently invented statistic, and someone asked me for the source. There wasn't one. There had never been one. It read exactly as convincingly as the true sentences around it, which is precisely the problem — fluency isn't accuracy, and these systems are equally fluent when they're wrong.

I over-engineered everything. After learning this stuff, I wrote 400-word prompts to ask what time zone a city was in. Most questions are just questions. The formula is for when output disappoints, not for everything.

What actually changed: I stopped thinking of it as talking to a computer and started thinking of it as briefing a smart contractor who's never met me, can't ask questions, and starts work the second I stop typing. Everything I'd need to tell that person, I tell the model. That reframe was worth more than every technique in this article.

Frequently Asked Questions

How long should a prompt be?

Long enough to remove ambiguity, no longer. Most useful prompts land between three and fifteen lines. The real test isn't word count: could a competent freelancer do this job correctly from your prompt alone, without asking a single follow-up? If yes, it's long enough. If no, the missing answers are what you should add.

Do longer prompts always give better results?

No. Length helps only when the extra words carry decisions — audience, purpose, format, things to avoid. It hurts when it adds vague praise like "engaging, powerful, world-class," because those describe a feeling and the model can't act on a feeling. Long prompts also make self-contradiction more likely, which is worse than saying less. Add words a stranger could check you against; delete words that only signal enthusiasm.

Does ChatGPT need different prompts from Claude?

The fundamentals transfer completely — role, context, audience, constraints, and format help everywhere. Differences are degree, not kind. In my testing Claude follows long structured instructions and negative constraints closely, while ChatGPT needs a firmer shove out of its default polished register. Both publish prompting guides worth reading, but you don't need a separate skill set per tool.

Should I use markdown or formatting inside prompts?

Yes, and it helps more than it should. Headings, bullets, and delimiters make instruction boundaries explicit, so your pasted content doesn't get read as an instruction and your constraints don't get skimmed. Label sections, fence pasted material with ---, keep constraints in a list rather than buried in prose. Costs nothing, measurably improves instruction-following on long prompts.

Can better prompts reduce hallucinations?

They reduce them; they don't eliminate them. Three habits genuinely help: give it the source rather than relying on its memory, explicitly permit "I don't know," and ask it to mark low-confidence claims. What no prompt can do is make a model reliably aware of its own knowledge limits — so for anything with consequences (a statistic, a citation, a medical or legal claim, a number in an exam answer), verify against a real source regardless of how good the prompt was.

Can prompts improve AI coding output?

Coding is where prompt quality pays off fastest, because code has an objective standard — it runs or it doesn't. What matters: language and version, environment, input/output shape, error handling, and what you've already tried. Pasting the actual error rather than describing it is the highest-value habit there is. Asking for the explanation alongside the fix turns it from a patch dispenser into something you learn from.

What is few-shot prompting?

Showing two or three examples of the input-to-output pattern you want before the real task. It's the most reliable way to communicate a style you'd struggle to describe. Three summaries you already like teach it faster than a paragraph of adjectives. Risk: it copies examples too literally — so vary them, and say explicitly that they demonstrate the pattern, not the content.

Does giving the AI a role actually work?

Yes, but not for the reason people think. Telling it it's an expert grants no expertise it lacked. What a role does is narrow the vocabulary, structure, and priorities it draws on, making output more consistent and context-appropriate. Useful roles are concrete: "a hiring manager who screens sixty resumes a week." Useless roles describe status: "a world-class genius." The model can copy behaviour, not status.

Why does AI give generic answers?

Because a generic question has a huge number of acceptable answers, and the model lands near the middle. Ask for a post about AI and a bland post about AI is a correct response — nothing you said ruled it out. Every detail you add deletes a large set of possible answers and pushes the output away from average and towards what you meant.

Is prompt engineering a real skill or just hype?

Real, and badly oversold. The hype version promises secret words and magic templates; that version is nonsense. The real version is unglamorous — the ability to specify a task clearly, which is the same skill that makes someone good at briefing a colleague or writing a bug report. An afternoon to learn the fundamentals, years to get genuinely good, mostly because the hard part is knowing what you want rather than how to phrase it.

Should I ask the AI to write my prompt for me?

It's a legitimate shortcut and I use it, with a caveat. Describing your task and asking for a detailed prompt, then editing the result, often beats a blank box. But the model doesn't know your audience, constraints, or taste — what you get is a competent skeleton with your specifics missing. Treat it as a draft to correct, never as finished, or you've just moved the vagueness one step back.

How do I stop AI writing from sounding like AI?

Ban the specific patterns rather than asking for a human tone — "human tone" isn't actionable. Name the words you never want, prohibit the habits you find hollow (opening on a rhetorical question, closing on an inspirational summary), demand one concrete detail only real experience would produce, and set a hard word limit. The limit does the most work, because padding is where the tells live. It'll still need your edit, but it starts far closer.

Do I need to rewrite a prompt five times every day?

No. The five rewrites here are a teaching device for a process that collapses into one good prompt once you've learned it. In daily use I write something like version five first time, because the decisions became habits. Deliberate rewriting is worth it when the task is new, when output disappoints and you don't know why, or when you're building something reusable. For a throwaway question, a throwaway prompt is fine.

Can I reuse the same prompt for different tasks?

Yes, and building a small library is probably the highest-return habit here. Write them as templates with obvious blanks, keep them in a plain text file rather than trusting chat history, and revise any template that produced something you had to heavily edit. Five or six refined prompts will cover most of what you actually do, and they improve every time you fix one.

What's the most common prompting mistake?

Assuming the model knows what's in your head. You know who it's for, why you're writing it, what happened last week, what you'd never say. None of that is in the prompt, and it can't ask, so it fills every gap with the most ordinary option available. Nearly every disappointing output I've produced traces back to something I knew and didn't share.

Key Takeaways

• The model isn't the bottleneck. Same tool, same session, six wildly different outputs. Only the prompt changed.
• Context and audience do most of the work. Everything after them is refinement. If you add two lines, add those.
• Generic in, generic out — and it's not a flaw. Average is the correct answer to an unspecified question.
• Structure creates depth. Telling it which section mattered most is what produced the best insight in the whole experiment.
• Ban words, don't request tones. "Human" is unactionable. "Never use 'leverage'" is checkable.
• Hard word limits kill clichés, because padding is where clichés live.
• Roles work by narrowing, not by granting expertise. Describe behaviour, never status.
• Fluent isn't accurate. These systems sound equally confident when wrong. Verify anything that matters.
• Don't over-engineer. Most questions are just questions. The formula is for when output disappoints.
• Save what works. A handful of refined templates beats a folder of downloaded ones.

Final Thoughts

Go back and read output one and output six again. Same model. Same afternoon. Same underlying facts about some invoices.

The gap between them wasn't intelligence, or a better subscription, or a secret phrase someone's selling in a course. It was that by version five I'd actually decided what I wanted — who it was for, what it needed to do, what it must never sound like — and version zero was me hoping the machine would decide for me.

That's the uncomfortable part, honestly. Most bad prompts aren't a writing problem. They're a thinking problem wearing a writing costume. You can't specify an audience you haven't considered or ban a cliché you haven't noticed. Getting good at this made me better at briefing people, too, which I did not expect and which says something about how vague most of us are by default.

So try it once, properly. Take the last request where AI let you down — the one that made you close the tab — and rebuild it. Add the context you skipped. Name the reader. Give it a shape. Ban five words. Then look at what comes back and tell me the tool was the problem.

The best AI users aren't the ones with the smartest AI. They're the ones who ask better questions — and the reason they ask better questions is that they did the thinking first.

Related reading on this site: Using AI for Coding Interview Prep, A Free AI Study Stack, and My AI Resume Got Rejected.