40 ready-made AI prompts — and what to do when they don't work

In short: a ready-made prompt is a template that already contains the role, task, context and format — you just fill in the square brackets. Below are 40 of them across six categories. But the list is the easy half. We spent July 2026 pulling data out of Udemy's internal API and Coursera's review pages, and found that the thing learners complain about most is the thing almost nobody teaches: your prompt didn't work — what do you change first? So the second half of this piece is debugging, common failure modes, and prompts you can run right here on the page with a button.

First, an honest problem with lists like this one

There are more "100 best prompts" lists online than there are people using them. We wanted to know what the biggest paid courses actually teach, so we went and got the numbers ourselves — from Udemy's internal API and from Coursera's public review pages, measured 17 July 2026.

The scale first. On Coursera, Vanderbilt's "Prompt Engineering for ChatGPT" has 698,444 enrolments. Vanderbilt's longer "Prompt Engineering Specialization" has 138,967. IBM's "Generative AI: Prompt Engineering Basics" has 654,256. "Google Prompting Essentials" has 365,425. On Udemy, RPATech's "Prompt Engineering with ChatGPT Masterclass" has 118,803 students, and Mike Wheeler's "Prompt and Context Engineering 101" has 84,942 at a 4.31 rating — the lowest in our sample of eight top courses.

Now the complaints. These are courses with prompt engineering in the title:

• "There is nothing teached about creating a good prompt. It is just an overview of types of prompts." — Geralt O., 01.07.2026, 2★ (Wheeler)
• "No specific guidance on prompt engineering… what to avoid while asking, how to organize your thoughts, how to give feedback to AI based on its answers etc." — Bharat Ram A., 05.06.2026, 1.5★ (Wheeler)
• "i thought it would go deeper in prompts and have more examples and sessions to master or enhance our current prompts." — Manuel L., 15.06.2026, 2★ (PE for Everyone)
• "Too much background on ChatGPT. Get me to how to prompt GPT" — Mark R., 16.09.2025, 1.5★ (Academind)

Four people, four different courses, one identical complaint. Not "boring", not "expensive" — but "you showed me a taxonomy of prompts instead of teaching me how to fix one". Bharat Ram A. even wrote out the syllabus he wanted and didn't get: what to avoid when asking, how to organise your thoughts, how to give the model feedback on its answer.

The conclusion we drew: the gap in this market isn't templates, it's debugging. A template answers "what do I type the first time". Debugging answers "what do I do when the first time didn't work" — which is roughly always. So this article is built accordingly: forty templates first, then the part those courses skip.

A caveat against ourselves. Our quotes come from the default review feed any visitor sees, not from a representative sample — Coursera's star filter runs client-side and its server HTML always returns the first page. So this is not a measure of how many people are unhappy; it's a read on what the unhappy ones are unhappy about. We did check the proportion, and it's small: negative reviews (≤3.5★) run 1.5–3.5% on Coursera, and noticeably higher on Udemy — 9.61% for "Generative AI for Beginners" and 10.36% for "The Complete AI Guide". Most people like these courses. It's just that the ones who don't, dislike the same thing.

Why the Run button under a prompt isn't decoration

There's a second finding, and it explains why prompting is so hard to teach on video. Here's a review of Vanderbilt's "Prompt Engineering for ChatGPT":

"you send your assignments and immediatly you got your results: 100% correct. I am still speechless. I put all that effort in and have no idea whether I my answer was correct or not." — Shinysheep, 21.09.2023, 1★

That isn't a lazy instructor. It's a structural limit: there is no model inside a Coursera lesson. There is exactly one way to check a prompt — run it and look at the answer. The platform can't do that, so it stamps an automatic "100% correct". The learner did the work and learned nothing about whether the work was any good.

The same limit produces the single most common complaint about video courses in our data — the instructor is reading the slide out loud. Verbatim: "why read straight from the slide? I can do that. This was not a helpful course at all" (Janie I., 02.07.2026, 1★). And harsher: "50% of this course is reading script like a robot from the slides… they are just reading text from the slides which you can also you from any good website" (Shashank T., 13.04.2026, 1★). One more, on a course with 118,803 students: "Written by AI, delivered by AI. The slides are way too crowded to be useful and it really doesn't help to have the bot read them out word-for-word" (Bradley M., 16.04.2026, 1★, RPATech).

On this page every <pre> block gets a Run button: you press it and see a live model answer without leaving the article. The prompts below are written to run as-is, with no substitutions — the example is already baked in. That's the thing a lesson with no model inside it structurally cannot give you.

How to use this list

• Fill in every bracket. An empty bracket is a hole the model plugs at random, and the answer turns generic.
• Add your own material. Almost any prompt gets stronger if you paste the source after it: the email, the draft, the data, the constraints.
• Don't stop at the first answer. "Shorter", "different tone", "replace the example" — 2–3 refinements usually get you there.
• A template is a hypothesis, not an incantation. It saves you the first minute. After that, the debugging section does the work.

If the "Role + Task + Context + Format" formula is new to you, start with what a prompt is and how to write one — then you'll be able to build your own, not just copy.

Work emails and correspondence (8)

The biggest time saver: not "write an email", but an email with a defined goal and tone.

• Ask for something: "You are an experienced business correspondent. Write an email to my colleague [name/role] asking for [what you need]. Context: deadline [date], my previous request went unanswered, the relationship is good and professional — no pressure. Under 120 words, polite, with a clear deadline and one question at the end."
• Soften a harsh email: "Here's my draft: [text]. Rewrite it so the meaning and the ask stay intact but the tone becomes calm and professional. In two bullets, explain what you changed and why."
• Polite decline: "I've been offered [what]. I'm declining because [reason]. Write a 3–4 sentence decline: no self-flagellating apologies, but leave the door open."
• Reply to an unhappy customer: "The customer writes: [complaint]. The facts on my side: [what actually happened]. Write a reply: acknowledge the problem, explain without excuses, propose a concrete next step with a date."
• Chasing an overdue task: "Write a third reminder about [task]. Tone: friendly, but now naming the consequence of the delay: [consequence]. Under 80 words."
• Meeting recap: "Here are my meeting notes: [text]. Build a recap email: decisions, owners, deadlines. As a separate list — open questions with no decision yet."
• Email in another language: "Translate this email into [language]: [text]. Keep the meaning but adapt the phrasing to that language's business norms rather than translating literally. Flag places where the original tone would read as rude."
• Triage an incoming email: "Here's a long email: [text]. Answer three questions: what do they want from me, by when, and what happens if I don't reply. One sentence each."

What actually does the work here isn't the role ("experienced business correspondent" contributes very little). It's two things: a stated length, and one behavioural constraint ("no pressure"). A behavioural constraint beats a role almost every time — remember that, because it's step three of the debugging list further down.

Writing and content (6)

• Cut to N words: "You are an editor who cuts filler. Shorten the text below to [N] words without losing a single fact or conclusion. Audience — [who], reading on a phone, short on time. Give the shortened text first, then a list of what you cut. Text: [paste yours]"
• Headlines: "Here's the text: [text]. Give 10 headline options: 3 neutral, 3 with a number, 3 as a question, 1 provocative. No clickbait — the promise must be delivered by the text."
• Outline before writing: "I'm writing [what] for [whom]. The reader's goal: [why they care]. Draft an outline of 5–7 sections with one sentence each on what belongs there. Don't write the text itself."
• Find the weak spots: "Read this as a sceptical reader: [text]. Name 5 places where you'd stop believing or get bored, and explain each. Don't rewrite."
• Adapt across channels: "Take this text: [text]. Make three versions: a 600-character post, a newsletter email, and a one-minute video script. One message, three shapes."
• Match my voice: "Here are 3 of my texts as a style sample: [texts]. Write a new one on [topic] hitting the same voice. Then list which style markers you reproduced."

"My own voice" is its own pain, and no template fixes it outright. From the OpenAI forum thread on writing without the AI flavour: "I've studied best practices to avoid it generating content that sounds like AI, but I'm not having any success." From Hacker News: "the email writing tools just seems to strip out my personal voice making me sound like I'm writing unsolicited marketing spam." Caveat: those are advanced users and only a couple of sources — a signal, not a measurement. But the signal is consistent, and it points somewhere useful: samples of your own writing inside the prompt do more than any adjective like "write vividly" ever will.

Learning and unpacking hard things (6)

• Explain from zero: "You are a patient teacher. Explain [topic] so that someone with no background gets it. First an everyday analogy, then a 150-word explanation of the substance, then three questions to check my understanding — and wait for my answers, don't answer for me."
• Debrief a mistake: "I solved it this way: [my solution]. The correct answer is: [answer]. Don't hand me the solution — show me which step I went wrong at and why that's a typical error."
• Study plan: "I want to learn [skill] in [timeframe], I can spend [N] hours a week, starting level [what]. Build a week-by-week plan: what to do, what to practise on, how to check the week landed."
• Socratic mode: "Don't explain [topic] to me directly. Ask one question at a time and lead me to the understanding. If I'm wrong, don't correct me — ask a guiding question."
• Summarise an article: "Here's the text: [paste]. Give me: 1) the main claim in one sentence, 2) the author's three arguments, 3) what the author asserted but never proved."
• Revision cards: "Make 15 question-answer cards on [topic] for a [level] learner. Answers no longer than two sentences. No guessing questions."

Note the "wait for my answers, don't answer for me". Without that clause the model will ask its three questions and immediately answer them itself — its job is to continue text, not to examine you. It's the same hole as the automatic "100% correct": with nothing checking you, you don't find out whether you understood.

Analysis and decisions (7)

• Pressure-test a decision: "You are an analyst who argues with me instead of agreeing. Pressure-test this decision — [what I'm about to do]. Goal [what], constraints [money/time/people], deadline [when]. Format: 3 strong arguments for, 3 against, the risk I've most likely missed, and one question whose answer would change the decision."
• Devil's advocate: "Here's my plan: [plan]. Find its weakest link and explain how it falls apart in reality. Don't soften it."
• Compare options: "Compare [option A] and [option B] for [task]. Format: a table — criterion, A, B, what matters most for me. Choose the criteria yourself and justify them."
• Read the data: "Here are the numbers: [data]. Name three conclusions that follow from them, and separately — conclusions that do NOT follow even though they look logical."
• Negotiation prep: "I'm negotiating with [who] about [what]. My position: [what]. Play the other side: what three objections will you raise and what do you actually want?"
• Break a task into steps: "Task: [what needs doing]. Break it into steps so the first takes no more than 15 minutes. Mark the step where people usually give up."
• What am I missing: "Decision: [what]. I've considered: [list]. Name 5 factors missing from my list that could change everything. For each, how to test it in a single day."

Career and money (5)

• Tailor a CV: "Here's my CV: [text]. Here's the job ad: [text]. Rewrite the experience section for it: same facts, emphasis on what's relevant. Flag what the CV is missing for this role."
• Interview prep: "Job ad: [description]. My background: [brief]. Ask me 8 questions they'll realistically ask, including two uncomfortable ones. One at a time, wait for my answer."
• Raise conversation: "I want to discuss a raise. My results this year: [list]. Build the case: what to say, in what order, and how to answer 'there's no budget right now'."
• Assess an offer: "Offer: [terms]. What questions should I ask before signing? What usually turns out to matter more than it seems up front?"
• Weekly plan: "Here are my tasks: [list]. Sort them by urgent/important, suggest what to drop entirely, and name the one task that yields the most."

Everyday life (8)

• Meals and groceries: "Build a week's menu: [how many people], budget [amount], no [allergens/foods], cooking under 30 minutes. Separately — a shopping list grouped by supermarket aisle."
• Spending review: "Here's my month of spending: [list]. Group it into categories, find the three where most leaks out, and suggest one realistic cut for each."
• Trip planning: "I'm going to [city] for [how many days], interests: [what]. Build a day-by-day route that respects geography so I'm not zigzagging. Flag what to book in advance."
• A difficult conversation: "I need to tell [who] about [what]. I'm worried they'll [reaction]. Help me find the wording: what to say first, what not to say at all."
• Instructions in plain language: "Here's a manual: [text]. Rewrite it step by step for someone doing this for the first time. Mark the step people most often get wrong."
• Gift ideas: "A gift for [who], budget [amount], they like [what], dislike [what]. Give 10 ideas: 5 conventional, 5 non-obvious. No clichés like a gift card."
• Rehearse a talk: "I'm speaking [where] for [how long] to [whom]. Topic: [what]. Give me a structure, an opening line and a closing line. Then ask 3 questions the room might raise."
• Contract check: "Here's a contract: [text]. Find the clauses that work against me and explain each in one plain sentence. Separately: what's missing that should be there. You're not a lawyer — flag where I need one."

Now the part that matters: your prompt failed — what do you change first?

This is the gap the whole article was written for. The four people quoted at the top paid money and walked away without this skill. Work the list top to bottom and don't skip.

• 1. Check whether the prompt contains a definition of "done". The most common cause of a mediocre answer is that the model has no idea what a good one looks like. "Write some copy about the product" has no criterion, so you get the average of all product copy ever written. "Write 150 words that would persuade someone who already said no once" has a criterion. If you can't articulate what separates a good answer from a bad one, the model certainly can't.
• 2. Add material, not adjectives. "Make it more interesting" changes nothing, because "interesting" is not an operation the model can perform. Three paragraphs of your own writing pasted in as a sample changes everything. The rule: to improve an answer, add input, not a verdict on the last one.
• 3. Replace prohibitions with prescriptions. "Don't be corporate" performs worse than "sentences under 12 words, active verbs, no noun stacks". A prohibition says where not to go; a prescription says where to go. Models are much better at doing a thing than at not doing a thing.
• 4. Cut the task apart. If one prompt asks the model to find, judge and rewrite at once, all three come out weak. Ask "name 5 weak spots" first, then "rewrite spot #3". The model doesn't get tired, but it still has one thread of reasoning per answer.
• 5. Ask it what it was missing. Literally: "Before you answer, ask me the 3 questions whose answers would most change your response." It is the cheapest way to find the hole in your own prompt — and it's precisely what Bharat Ram A. asked for and didn't get: how to give the model feedback on its answer.
• 6. Only then touch the role and the wording. "You are an experienced marketer" is a few-percent tweak. People start there because it's the visible part, and they get stuck there because it's the weak part.

Try it on a live example. The prompt below is self-contained — hit Run, nothing to fill in: a deliberately bad prompt is already inside it, waiting to be fixed.

You are a teacher of prompt debugging. Here is a prompt written by a beginner:

"Write a good social media post about our app. Make it interesting and don't be boring."

Do four things:
1. Name exactly three reasons this prompt will produce a generic answer.
2. For each reason, state the edit that fixes it — the edit must be an operation, not a verdict.
3. Rewrite the prompt in fixed form, inventing the missing context yourself, and mark everything you invented with the word ASSUMPTION.
4. Finish with one sentence: which of the three edits would give the biggest gain, and why.

Second experiment: specificity beats politeness and beats roles. This one makes the model demonstrate the difference on a single task.

Complete three assignments in a row and show all three answers in full.

Assignment A. Respond to this prompt: "Write an email about moving a deadline."
Assignment B. Respond to this prompt: "You are an experienced business correspondent. Write a professional email about moving a deadline."
Assignment C. Respond to this prompt: "Write an email to a colleague: I'm asking to move the report deadline from Friday to Wednesday next week, because our contractor delivered the data late. The relationship is good, no pressure, and this is my first request to move it. Under 90 words, one question at the end, apologise no more than once."

After the three answers, build a table: which properties appeared in B compared to A, and in C compared to B. In the last row, answer this question: which added more — the role, or the context?

Third move, the most underrated one. Don't ask the model to answer. Ask it to interrogate you first.

My situation: I have one week to choose between two job offers and I don't want to regret it a year from now.

Do not answer and do not give advice yet. Instead:
1. Ask me exactly 5 questions — the ones whose answers would most change your recommendation. Order them by how much they matter.
2. For each question, explain in one sentence what part of your recommendation depends on it.
3. Then state what your recommendation would be if I answered none of the questions, and why that version should not be trusted.

Wait for my answers before going any further.

Notice what all three have in common: none of them ask for a finished artefact. They ask the model to expose its own reasoning, its own assumptions, or the difference between two inputs. That's what a checkable exercise looks like — and it's why a prompt you can press Run on teaches more in thirty seconds than a slide about prompt taxonomy does in twenty minutes.

Five mistakes that kill a template

• Leaving the brackets unfilled. It sounds silly, but it's cause number one. The model won't ask — it will make something up.
• Stuffing three tasks into one prompt. "Analyse, rewrite and suggest headlines" yields three mediocre chunks instead of one good one.
• Demanding "no filler" without stating a length. Length is the one constraint a model follows literally. Use it.
• Believing the first answer. Every answer is a draft. Especially the confident-sounding one.
• Collecting templates instead of collecting edits. Forty of someone else's templates are worth less than three of your own, debugged against your own work.

What the same thing costs as a course

Since we were already inside the pricing data, here it is, measured 17 July 2026. The common belief that "Udemy is always 90% off" doesn't survive contact with their own API, which returns "saving_price": 0.0, "has_discount_saving": false, "discount_percent": 0. There is no discount. Eleven of the twelve top courses sit at €19.99; one (Complete AI Guide) at €24.99. Udemy Personal Plan is €20.00/month, €10.00/month on promo. Coursera Plus is €50/month or €343/year, with a 14-day refund window.

And one detail worth knowing before you buy: Vanderbilt's "Prompt Engineering Specialization" requires a paid ChatGPT+ subscription to complete the assignments. So the practice costs you twice — once for the platform, once for the model. We're not quoting the exact ChatGPT Plus price here: the primary source is closed (every OpenAI domain returned 403 to our crawler), and second-hand numbers don't belong in an article that's arguing for checkable claims.

What to do next

Don't try to memorise forty templates. Pick three that match real tasks and run them this week. When an answer comes back wrong — and it will — walk the six debugging steps top to bottom and write down which one fixed it. In a month you won't have a list of someone else's prompts; you'll have a skill that survives the next model release, which the list won't.

One caveat: every answer to these prompts is a draft, not the truth. A model can confidently invent a fact, a number or a link — why that happens is covered in AI hallucinations. Which data is safe to paste into a chat at all is in our piece on privacy when working with AI. The underlying formula is unpacked in what a prompt is. And if you want a prompt that doesn't just answer but carries out a multi-step task on its own, that's the territory of AI agents.