50 Best ChatGPT Prompts for Every Use Case

Most ChatGPT prompt lists are sorted by job: prompts for marketing, prompts for students, prompts for coding. That is useful while your job matches the list and useless the moment it does not, because what makes a prompt work is not its topic but its structure.

The fifty below are sorted by technique instead: assigning a role, controlling the output format, forcing the reasoning into view, teaching by example, and refining what comes back. Learn the five patterns and you can write your own prompt for anything, which is worth more than memorising a list. ChatGPT responds to all five, and so does every comparable model.

Each prompt has bracketed parts to replace. Copy it, swap the brackets, keep the structure. Run them in ChatGPT directly or in Chat Smith, which is useful here mainly because it lets you send the same prompt to a second model and compare without retyping anything.

A request names a task. A prompt names the task and then removes the decisions you do not want the model making on your behalf. There are five reliable ways to do that, and the sections below are built around them.

• Give it a role and a reader - who is answering, and who is listening. This narrows vocabulary, assumptions and register all at once.
• Specify the output - a table, a word count, valid JSON, one question per turn. Anything unspecified comes back as prose.
• Make the reasoning visible - ask for steps, assumptions and confidence rather than conclusions you cannot check.
• Teach by example - two or three examples of what you want beat a paragraph describing it. This is the technique people reach for last.
• Refine rather than restart - when output misses, diagnose the prompt instead of rewriting it blind.

A request looks like this:

Give me some marketing ideas for my product.

The same thing as a prompt:

Act as a growth lead who has launched three products in [category] and is sceptical of paid acquisition. My product is [description] and my main constraint is [budget or team size]. Give me eight ideas in a table with columns for idea, why it suits this constraint, the first test to run, and how I would know it had failed. One sentence per cell, no preamble.

Same request, completely different return. Every prompt below is built that way, which is why they are longer than the one-liners you are used to.

Assigning a role is the cheapest improvement available. Naming a specific reader narrows it again. For applying this to drafting in particular, our ChatGPT prompts for writing go deeper.

1. Expert with a stated limitation

Act as a [profession] with fifteen years of experience who is sceptical of trends. I will describe [situation]. Give me your assessment in under 300 words, and state plainly the one thing you would need to know before advising further.

2. Two experts disagreeing

Simulate a short exchange between two specialists who genuinely disagree about [question]: one favouring [position A], one favouring [position B]. Six turns, no resolution, neither may concede. Then list what they do agree on.

3. The specific reader

Explain [topic] to a [specific role, for example a hospital procurement manager] who has thirty seconds and one decision to make. No background, no history, only what bears on that decision.

4. The hostile reviewer

Act as a reviewer who wants to reject this. Read the material below and give me the three objections most likely to sink it, ordered by severity. Do not soften anything and do not suggest fixes yet.

5. Beginner and expert side by side

Explain [concept] twice: once for someone meeting it for the first time, once for someone who works in the field and wants the nuance. Label both. The second version must not repeat the first.

6. Devil's advocate on my own plan

Here is my plan. Argue against it as convincingly as you can, using only reasoning I have not already addressed in it. Do not be balanced. I will supply the defence myself.

7. The translator between functions

Act as someone who has worked in both engineering and sales. Rewrite the technical explanation below so a sales team can use it in a client conversation, keeping every claim accurate and flagging anything they should not promise.

8. Interviewer, not answerer

Act as an interviewer preparing me for [role]. Ask me one question at a time, wait for my answer, then follow up on the weakest part of it. Do not move on until the answer holds, and do not give me model answers.

9. A role with a constraint on tone

Act as a mentor who is warm but does not flatter. Read my draft below and tell me what is genuinely working and what is not, in roughly equal measure, without any encouragement that is not earned.

10. Audience objection map

I am presenting [proposal] to [specific audience]. List the six objections that audience is most likely to raise, ranked by likelihood, and for each say whether it is a real problem or a communication problem.

Give the role a limitation, not just a title. A sceptical expert of fifteen years produces better output than an expert, because scepticism is a behaviour while expertise is only a label.

The most common reason output is unusable is that nobody said what shape it should be. Say table, say 100 words, say valid JSON, say no preamble. Structured output matters most in technical work, and our ChatGPT prompts for SEO cover the recurring formats in that area.

11. Table with defined columns

Compare [three options] in a table with exactly these columns: option, best for, biggest limitation, cost signal, and what would make you rule it out. One short sentence per cell. No prose before or after the table.

12. Hard word count

Summarise the text below in exactly 100 words. Count them and state the count at the end. If it will not fit, cut a claim rather than compressing sentences into something unreadable.

13. Banned words

Write a product description for [product] in under 120 words without using the words innovative, seamless, revolutionary, leverage, robust or game-changing. If you find yourself reaching for a synonym of any of those, be more specific instead.

14. Structured JSON output

Extract the following from the text below and return only valid JSON with no commentary: an array of objects with the keys name, role, organisation and quote. Use null where a field is absent rather than guessing at it.

15. Fixed section skeleton

Write a one-page brief on [topic] using exactly these headings: Situation, What Changed, Options, Recommendation, What We Need to Decide. Under 60 words per section. No introduction and no conclusion.

16. Ranked list with reasoning

Give me eight options for [problem], ranked, and for each state the single reason it holds that position. Then tell me which two you would drop entirely and why.

17. Two versions, opposite strategies

Write two versions of [message]: one direct and short, one that builds context first. Label each with what it optimises for and what it risks. Do not tell me which to send.

18. One question per turn

I want to work through [problem] with you. Ask me one question at a time and wait for my answer before the next. Do not summarise, do not offer solutions, and do not ask multi-part questions.

19. Progressive detail

Explain [topic] three times at increasing depth: one sentence, one paragraph, one page. Each level must be independently readable rather than a continuation of the one before it.

20. Explicit refusal instruction

Answer the question below using only the document I have attached. If the document does not contain the answer, say so plainly and stop. Do not fill the gap from general knowledge.

If you find yourself reformatting output by hand, that is a prompt problem rather than a model problem. Put the format in the prompt and it stops happening.

Asking for a conclusion gets you a conclusion. Asking for the steps, the assumptions and the confidence level gets you something you can actually check. This is the group that matters when the answer has consequences, and our ChatGPT prompts for data analysis apply the same idea to figures specifically.

21. Show the working

Work through [problem] step by step, stating each assumption as you use it. At the end, tell me which step you are least confident about and what would change the answer.

22. Decomposition first

Before solving anything, break [complex problem] into its independent sub-problems and tell me which one to address first and why. Do not attempt a solution in this response.

23. Assumption audit

Here is a plan. List every assumption it depends on, ordered from most to least load-bearing. For the top three, say what happens to the plan if each one turns out to be wrong.

24. Working backwards from failure

Assume [project] failed badly twelve months from now. Write the post-mortem: the three most likely causes, the early warning signs each would have produced, and what could be monitored now to catch them.

25. Steelman, then rebut

Build the strongest case for [position I disagree with], using only arguments its real proponents make. Then give the two strongest objections to that case. Do not indicate which side you find more persuasive.

26. Estimate with stated inputs

Estimate [quantity] and show your reasoning as a chain of stated inputs and multipliers. Flag which input the answer is most sensitive to, and give a plausible range rather than a single figure.

27. Evidence versus assertion

Read the argument below and classify every claim as supported by evidence given, asserted without support, or assumed implicitly. Do not comment on whether the conclusion is right.

28. Second-order effects

If [change] happens, list the first-order consequences, then the second-order consequences those would produce. Stop at second order, and mark any that would only occur under specific conditions.

29. Comparison on my criteria

I care about [criterion A] most, [criterion B] second, and not at all about [criterion C]. Compare [options] against that weighting specifically, and tell me if my weighting makes any option obviously wrong for me.

30. Anomaly hunting

Review the data below and identify anomalies: outliers, impossible values, suspicious duplicates, and any column whose distribution suggests a data-entry problem rather than a real pattern. Explain your reasoning for each.

Ask for the confidence level as a separate line. A model that has to label a claim low-confidence tends to hedge in the right places instead of hedging uniformly.

Two examples of what you want beat a paragraph describing it. This is the technique people reach for last and should reach for first. It transfers to visual work as well, which is the same logic behind our ChatGPT prompts for image generation.

31. Few-shot pattern

Here are three examples of the output I want, with their inputs. Study the pattern, state in one sentence the rule you have inferred, then apply it to the four new inputs below. Do not deviate from the pattern in order to improve it.

32. Voice analysis, then reuse

Here is a 300-word sample of my writing. Analyse it for sentence-length variance, vocabulary level, use of contractions and how it handles transitions. Then write 200 new words on [topic] matching those features, and list which features you matched.

33. The anti-example

Here is an example of output I do not want, and why. Write three alternatives that avoid that specific failure, and after each one note which part of the failure you were addressing.

34. Format by example

Copy the exact structure of the document below - same section order, same relative lengths, same level of formality - but for [different subject]. Do not improve the structure.

35. Calibrating with a rubric

Rate the five drafts below from 1 to 5 on clarity using this rubric: [rubric]. Then explain the difference between your 3s and your 4s specifically, so I can see where you drew the line.

36. Contrast pair

Write the same paragraph twice: once as a writer who trusts the reader to keep up, once as a writer who over-explains. Same content, same length. Label them.

37. Style transfer with content locked

Rewrite the passage below in the register of [genre or document type]. Every factual claim must survive unchanged. Afterwards, tell me which claims were hardest to preserve in the new register.

38. Consistency across a set

Here is one item written exactly the way I want. Produce nine more for the items listed below, matching its length, structure and level of detail. Flag any where matching the format forced you to leave out something important.

39. Deriving conventions from what worked

Here are two prompts that produced results I liked. Identify what they have in common in framing, level of detail and specification, then write five new prompts for [subject] using those same conventions.

40. Reverse-engineering a good output

Here is a piece of work I admire. Write the prompt most likely to have produced it, then explain which parts of that prompt are doing the heavy lifting.

Include one anti-example. Showing what you do not want is often faster than describing what you do, and it is the only reliable way to communicate tone.

The instinct when output misses is to rewrite the whole prompt. Almost always the fix is one clause. These ten are for finding which clause. They matter most in technical iteration, which is why our ChatGPT prompts for coding lean on the same pattern.

41. Ask what was ambiguous

The output you just gave me was not what I wanted. Rather than trying again, tell me which parts of my instruction were ambiguous and what you assumed. Then ask me the two questions that would resolve it.

42. Self-critique before delivery

Draft [deliverable], then critique your own draft as a hostile reviewer would, then give me only the revised version plus a two-line note on what you changed.

43. Confidence flagging

Answer the question below, then mark each claim as high, medium or low confidence. For every low-confidence claim, say what you would need in order to verify it.

44. One change at a time

Here is your previous output. Change only the opening paragraph and keep everything else identical. Then stop, so I can see the effect of that single change before we go further.

45. Deliberate over-generation

Give me fifteen options for [thing], including at least four you consider weak. Do not filter for quality. Then, in a separate list, tell me which four you think are weak and why.

46. The subtraction pass

Take the text below and cut 30 per cent without losing any argument. Show me what you removed as a separate list so I can veto individual cuts.

47. Testing my own reasoning

I concluded [X] from [evidence]. Tell me whether that inference holds, what alternative conclusion the same evidence would support, and what additional evidence would distinguish between the two.

48. Failure-mode check

Before I use the output below, tell me the three ways it is most likely to be wrong or to fail in practice, and what I should check first.

49. Iterating on the prompt itself

Here is a prompt I wrote and the disappointing output it produced. Diagnose the prompt rather than the output: what is missing, what is ambiguous, what is over-specified. Then give me a revised prompt without running it.

50. A stopping condition

Continue improving the draft below, but stop as soon as further changes would be a matter of taste rather than quality. Tell me when you reach that point and what the remaining choices are.

Never regenerate blind. Asking which part of the instruction was ambiguous costs one turn and usually converts three more turns of guessing into a single clause.

The same prompt sent to two models produces visibly different output, which is the whole argument for having more than one available. The reasoning prompts above are the best place to see it: run two of them through GPT-5.6 Sol and GPT-5 Nano and the difference in how each shows its working is immediate. The full model list is worth going through once so you know what you have to hand.

One habit worth forming, and it is prompt 41 on this list: when output misses, do not rewrite the prompt. Ask which part of your instruction was ambiguous. The model will usually tell you, and the fix is normally a single clause. That one move will save you more time than any individual prompt in this article.