These prompts are guidance written for this page. We haven't published saved outputs for them, and results will vary with your draft.
Why vague prompts don't work
OpenAI's own prompt engineering best practices make the case. The guide says to be specific about the desired context, outcome, length, format and style, to cut "fluffy" descriptions (its example swaps "fairly short" for "a 3 to 5 sentence paragraph"), and to say what to do instead of only what not to do. "Sound more human" fails all three tests: no reader, no length, no instruction about what human means.
The prompts below follow those rules. Each one names a target, sets a length or a format, and tells ChatGPT what to do rather than only what to avoid.
Prompt 1: the plain language rewrite
``` Rewrite the text below for a reader who is busy and not an expert. Use short, common words. Use active voice: say who does what. Keep every fact, name, number and date exactly as written. Keep it about the same length or shorter.
Text: """ [paste your draft here] """ ```
What it targets: passive sentences, long words and buried verbs. It borrows from the US government's plain language guidance on writing for understanding, which recommends shorter words, active voice and present tense.
What it tends not to fix: sameness of rhythm. A plain draft can still march along in sentences of identical length. It also won't add the specific detail that makes writing sound like a person; there's nothing in the prompt to add.
The triple quotes around the text follow another tip in OpenAI's guide: put instructions first and separate them from the text with ### or """.
Prompt 2: cut the stock phrases
``` Edit the text below. Remove or replace these phrases wherever they appear: "In today's world", "Ultimately", "Overall", "I hope this helps", "serves as", "stands as". Replace "serves as" and "stands as" with "is" where that keeps the meaning. Delete any opening sentence that only sets the scene. Delete any closing sentence that only repeats an earlier point. Change nothing else.
Text: """ [paste your draft here] """ ```
What it targets: the vocabulary and set phrases readers spot fastest. Wikipedia's editors list many of these in their field guide to signs of AI writing, including "serves as" in place of "is" and leftover chat phrases like "I hope this helps".
What it tends not to fix: new stock phrases. Ask ChatGPT to drop one and it may reach for another. Read the result for fresh filler and add those to the list next time. "Change nothing else" helps keep the edit narrow.
Prompt 3: rhythm and sentence length
``` Rewrite the text below so the sentence lengths vary. Follow at least one long sentence with a short one. No more than two sentences in a row of similar length. Split any sentence over 30 words. Do not add new ideas, examples or facts.
Text: """ [paste your draft here] """ ```
What it targets: the flat, even rhythm that makes a ChatGPT draft feel machine-made even when every word is fine.
What it tends not to fix: hedging and vague words. A paragraph can have a lively rhythm and still say "may potentially help to improve outcomes". Pair this prompt with Prompt 4, or fix hedges by hand.
Prompt 4: make the claims and drop the hedges
``` Edit the text below to be direct. Remove stacked hedges such as "may potentially", "could help to" and "it can be argued". Where a claim is certain, state it plainly. Where a claim is uncertain, keep one hedge and add the reason in a few words. Replace vague words (enhance, valuable, key, foster) with a concrete statement, or mark the spot with [DETAIL NEEDED] if you don't know one.
Text: """ [paste your draft here] """ ```
What it targets: soft, committee-style sentences that commit to nothing.
What it tends not to fix: missing facts. ChatGPT can't know the number, name or result your sentence needs. That's why this prompt asks it to flag gaps with [DETAIL NEEDED] instead of inventing something. Fill those in yourself, and delete the markers before you send anything.
Prompt 5: write it in my voice
``` Here are two paragraphs I wrote myself: """ [paste a sample of your own writing] """ Notice my sentence length, word choice, contractions and how formal I am. Now rewrite the text below to match that voice. Keep every fact the same. Do not copy sentences from my sample.
Text: """ [paste your draft here] """ ```
What it targets: generic voice. It follows OpenAI's advice to show the desired output through examples rather than only describe it.
What it tends not to fix: content that isn't yours. If the draft makes points you wouldn't make, matching your voice won't help. Also check that it hasn't lifted phrases from your sample into places they don't belong.
Prompt 6: draft it right the first time
Write a [email / cover letter / paragraph] for [who the reader is].
Purpose: [what you want the reader to do or know].
Use these notes, in this order: [your bullet points].
Length: [number] sentences.
Write like a person talking to one reader: contractions are fine,
short sentences are fine, no opening line that only sets the scene,
no closing summary, no lists of three unless there are exactly three things.
What it targets: the problem at its source. A draft built on your notes, for a named reader, at a set length, starts closer to your own writing than any rewrite of a generic draft.
What it tends not to fix: accuracy. ChatGPT will smooth over a gap in your notes rather than ask about it. Read the draft against your notes line by line.
Which prompt for which problem?
| What bothers you about the draft | Start with | Add if needed |
|---|---|---|
| Stiff, formal, hard to read | Prompt 1 | Prompt 3 |
| Obvious filler phrases | Prompt 2 | Prompt 4 |
| Every sentence feels the same length | Prompt 3 | Prompt 2 |
| Soft and vague, commits to nothing | Prompt 4 | Prompt 1 |
| Correct but not you | Prompt 5 | Prompt 3 |
| You haven't drafted yet | Prompt 6 | Prompt 2 after |
For emails specifically, our ChatGPT email humanizer page covers the phrases that give away a ChatGPT email and how to handle sign offs.
What prompts can't fix
A prompt changes what ChatGPT writes. It can't change three things.
It can't check facts. ChatGPT will keep a wrong date as faithfully as a right one, and may introduce errors in a rewrite. Read every name, number and quote against your original.
It can't judge its own result. Asking ChatGPT "does this sound human?" gets an opinion, not a measurement. Our guide on whether ChatGPT can humanize its own text explains why, using OpenAI's own help pages.
It can't decide what's allowed. If your school or employer has AI rules, a better prompt doesn't change them. Check, and disclose when asked.
When prompting stops paying off, edit by hand. Our list of nine fixes that make ChatGPT sound more human gives a before and after for each edit, and most take seconds. Or paste the draft into our free rewrite box, choose Casual, Professional or Academic, and make the final pass yourself.
Combining a prompt with a rewrite pass
The two work well in sequence:
- Draft with Prompt 6, using your own notes.
- Run Prompt 2 to strip stock phrases.
- Put the draft through a rewrite pass, such as our box, for rhythm and tone. It takes 20 to 1,000 words per run with no account.
- Read it out loud and fix any line you wouldn't say.
- Check facts against your notes.
Sources
- Best practices for prompt engineering with the OpenAI API, OpenAI Help Center, read November 8, 2026
- Writing for understanding, Digital.gov plain language guide series, US General Services Administration, read November 8, 2026
- Signs of AI writing, Wikipedia project page maintained by Wikipedia editors, read November 8, 2026
About the author
The ChatGPT Humanizer editors write and review the guides, run the tool comparisons to the published method, and look after the corrections inbox.
