50 ChatGPT Prompts for Writing to Write Better

The mistake most writers make with ChatGPT is asking it to write. It will, fluently, and you end up editing something that sounds like nobody. The prompts that actually help ask it to do the jobs around the writing instead: structure, diagnosis, compression, and arguing against you.

The fifty below are organised by where you are in the process: planning, drafting, editing, voice, and getting real feedback. That way you can go straight to the stage you are stuck at rather than reading a list. They work in ChatGPT and in any comparable model.

One thing to be clear about: these are prompts for getting help with writing, not prompts that give you something to write about. If you want the second kind, our short fiction prompts are a different set entirely. For prompting technique across every other kind of task, see the main ChatGPT prompts guide.

Four habits separate a prompt that helps your draft from one that replaces it.

• Ask for diagnosis before prose - what is weak and why, ranked. A model that rewrites first hands you fluency and hides the problem underneath it.
• Give it your words to work on - paste the draft. Prompts that start from nothing produce generic output; prompts that start from your paragraph produce edits.
• Constrain length precisely - exactly 400 words, under 80, cut 30 per cent. Vague requests for brevity produce vague results.
• Say what must survive - every factual claim, the paragraph count, my voice. Naming what has to be preserved is what stops a rewrite becoming a replacement.

A request looks like this:

Can you improve my blog post?

The same ask as a prompt:

Read the draft below as an editor who wants to cut it, not improve it. Name the three weakest paragraphs, say precisely why each is weak, and rank them by how much fixing would improve the whole piece. Do not rewrite anything, do not compliment anything, and do not suggest additions.

Same draft, completely different return. The first gets you a rewrite you did not ask for. The second gets you a list you can act on.

The blank page is a structure problem more often than an inspiration problem. These ten produce a shape you can write into.

1. Commissioning outline

Act as an editor commissioning a [word count] piece on [topic] for [specific reader]. Give me a section outline with an H2 for each section, one line stating what that section must accomplish, and a note on what to leave out. Flag any section where I will need a source.

2. Two structures, one subject

Outline [topic] twice using different organising principles: once chronologically, once by problem and solution. Same subject, same length. Then tell me which suits a reader who does not yet know they have the problem.

3. The reader's questions

I am writing about [topic] for [reader]. List the twelve questions that reader will arrive with, ordered by how early they need answering. Mark any I can safely leave unanswered.

4. Scope check

Here is my working title and a one-paragraph summary. Tell me whether this is one article or three, and if three, where the natural seams are.

5. Angle generation

Give me eight angles on [topic], each in one sentence, and for each say who it excludes. I want to see the trade-off, not just the options.

6. The argument in one sentence

Read my notes below and tell me what argument they are actually making, in one sentence. If they are making two, say so and tell me which is better supported.

7. Evidence gap map

Here is my outline. For each section, tell me what kind of evidence it needs - data, example, quotation, first-hand account - and which sections currently have none.

8. Opening candidates

Give me four possible opening moves for [topic]: a concrete scene, a surprising figure, a direct claim, and a question. Under 60 words each. Then tell me which is hardest to follow well.

9. What has already been said

For [topic], list the five points almost every article makes. I want to know what to skip, not what to include.

10. Ending before beginning

Given the argument below, write three possible closing paragraphs: one that resolves, one that opens outward, one that returns to the opening image. Under 80 words each.

Do not accept the first outline. Ask for a second built on a different organising principle, and choosing between them will teach you what the piece is actually about.

Use these when you are mid-draft and stuck, not when you want the draft written for you. The distinction matters: an AI paragraph generator can widen a section you have already chosen, but nothing can choose it for you.

11. Next move only

Here is where my draft stops. Do not continue it. Tell me the three things that could plausibly come next, and what each would commit the rest of the piece to.

12. Widen one section

Expand only the section marked below, keeping my voice and adding no new claims. Roughly double its length using detail and example rather than assertion.

13. Bridge the gap

I have these two paragraphs and no transition between them. Give me three possible bridging sentences, each making a different logical relationship explicit.

14. The weak middle

My draft has a strong opening and ending and a weak middle. Here it is. Tell me what the middle is supposed to be doing that it currently is not.

15. Concrete replacement

Here is a paragraph that is too abstract. List every abstraction in it and suggest a concrete image, example or number that could replace each. Do not rewrite the paragraph.

16. Talking it out

I am going to explain my argument to you badly, in conversation. Ask me one question at a time about the parts that do not follow. Do not summarise and do not write anything for me.

17. The missing step

Read the argument below and find the step I have skipped - the thing I know but have not said on the page. Quote where it should go.

18. Example generation

I need three examples to illustrate [claim]: one everyday, one professional, one historical. One sentence each. Then tell me if any of them do not actually illustrate the claim.

19. The concession paragraph

Write the paragraph where I concede the strongest counter-argument to [thesis], in a way that strengthens rather than weakens my case. Under 100 words, and do not resolve it too neatly.

20. Word target with substance

Continue the draft below by exactly 250 words, adding one new supported point rather than elaborating what is already there. Match the existing sentence rhythm.

Expand

Never paste a generated paragraph in unedited. Its job is to show you what the next move could be, and you will usually see a better one once it is on the page.

This is where a model genuinely earns its place. Compression, consistency checking and jargon auditing are mechanical work it does faster than you and without ego. Follow up with an AI sentence rewriter pass for rhythm and an AI grammar checker pass for mechanics - separately, not together.

21. The hostile read

Read the draft below as an editor who wants to cut it. Name the three weakest paragraphs, say why each is weak, and rank them by how much fixing would improve the whole. Do not rewrite anything.

22. Cut with a veto list

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

23. Exact length

Reduce the text below to exactly 400 words. Count them and state the count. If it will not fit, cut a claim rather than compressing sentences into something unreadable.

24. Structure map

Map the argument in the draft below as a numbered list of claims, marking each as supported, asserted, or assumed. Do not comment on style.

25. Consistency check

Read the whole document and list every internal inconsistency: contradicted facts, terms used differently in different places, numbers that do not reconcile, claims made and then dropped. Cite the location of each.

26. Jargon audit

List every phrase in the copy below that a reader outside the field would not understand, or would read differently than intended. Suggest a plainer alternative for each. Do not rewrite the whole thing.

27. Sentence-level flab

Find every sentence in the text below that could lose a third of its words without losing meaning. Show the original and the tightened version side by side. Leave all other sentences alone.

28. Redundancy hunt

Identify anywhere in the draft below that makes the same point twice in different words. Tell me which instance is stronger and why.

29. The opening and closing test

Delete the first paragraph and the last line of the draft below and show me what remains. Then tell me whether the piece lost anything or improved.

30. Reading-order check

Read the draft below and tell me whether any paragraph depends on information that only appears later. List each one with where that information currently sits.

Expand

Always ask for the cut list. A model that shows you what it removed lets you veto individual losses. One that just returns a shorter version has made those decisions for you, silently.

Tone requests fail when they name a label. Professional, engaging and conversational mean nothing operational. These ten specify behaviours instead: banned phrases, sentence-length targets, a sample to match.

31. Voice analysis, then match

Here is 300 words 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.

32. Banned phrases

Rewrite the passage below without using the words leverage, seamless, robust, innovative or delve, and without the construction it is not just X, it is Y. If you reach for a synonym, be more specific instead.

33. Register shift, content locked

Rewrite the passage below from formal to conversational. Keep every factual claim, keep the paragraph count identical, and add no humour or exclamation marks. Show before and after side by side.

34. Sentence-length prescription

Rewrite the paragraph below so that no sentence exceeds 20 words and at least three are under 8. Same content, same order.

35. Hedge removal

Find every hedge in the text below - arguably, somewhat, it could be said, generally - and either delete it or replace it with a specific qualification. Show me the list before you change anything.

36. Drift detection

Here are five sections of one piece. Tell me which ones drift out of the voice established in the first, and quote the specific sentences where it happens.

37. 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.

38. Formality calibration

Rate the formality of the text below from 1 to 10 and explain what specifically sets that level. Then produce a version two points lower without changing any claim.

39. Rhythm read

Tell me where the rhythm stumbles in the passage below: sentences that start the same way in sequence, clauses that pile up, paragraphs that all run to the same length.

40. The anti-example

Here is a piece of writing whose voice I do not want, and why. Write 150 words on [topic] that avoids that specific failure, and note which part of the failure you were guarding against.

Give it a sample of your own writing rather than an adjective. Three hundred words of you beats any description of you.

The most valuable thing a model does for a writer is disagree. These ten are for getting a real objection before a reader gives you one.

41. The unaddressed objection

Here is my thesis and draft. Give me the three strongest objections a well-informed sceptic would raise that my draft does not already address. Their strongest form, not straw versions.

42. Reader confusion points

Read the draft below as a reader who is intelligent but new to the subject. Tell me the first three places you would have to reread a sentence, and what confused you at each.

43. The reader who disagrees

Read the draft below as someone who arrived disagreeing with the conclusion. Tell me the earliest point at which you would stop reading, and why.

44. Claim classification

List every factual claim in the draft below and mark each as sourced in the text, commonly known, or needing verification. Do not check them, just classify them.

45. Promise audit

Read my headline and opening paragraph, then read the rest. Tell me what the opening promises and whether the piece delivers it. Be specific about any promise left unpaid.

46. The so-what test

For each section of the draft below, tell me in one line why a reader should care. If you cannot for a given section, say so - that is the answer I need.

47. Title against reality

Here are six possible titles and my draft. Tell me which titles the draft actually earns and which would be overselling it.

48. Confidence flagging

Mark each of the main claims in the draft below as high, medium or low confidence, based only on the support given in the text. Then tell me which low-confidence claim is carrying the most weight.

49. What this adds

Here is my draft on [topic]. Tell me in one sentence what it says that the standard treatment of this subject does not. If the answer is nothing, say so plainly.

50. Diagnose my prompt

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

Ask specifically for the objection you have not already answered. Without that constraint you get the three objections your draft already handles, which feels reassuring and tells you nothing.

For prose the model matters more than for most tasks, because the failure mode is so consistent: models over-write, reaching for the ornate phrase and the tidy summary. Claude Sonnet 5 tends to hold restraint when a prompt asks for restraint, while GPT-5.2 is quicker at generating structural options - six possible openings, three ways to cut a section. Sending the same paragraph to both takes a minute and settles the question faster than a third rewrite. The other AI writing tools handle the mechanical passes.

One habit worth forming: keep the prompts that worked in a file, with a note on what each one fixed. The diagnosis prompts especially - the hostile read, the structure map, the cut list - do the same job on every draft you will ever write, and rebuilding them from memory each time is the quietest waste in this whole workflow.