A grammar-check prompt works when it limits the edit, names the errors to look for, puts your text inside clear delimiters and fixes the output format. The eight prompts below follow guidance from OpenAI's and Anthropic's prompting docs and the error categories in Purdue's Online Writing Lab. They are original prompts, and we checked their logic on a seeded example, but we have not benchmarked any model's output, so verify every change before you accept it.
What the prompting guides recommend
OpenAI's prompt engineering guide recommends precise instructions, delimiters such as Markdown headers or XML tags to mark where reference content starts and stops, a few input and output examples, and an explicit output format. Anthropic's guide adds: be specific about the output format and constraints, give the reason behind an instruction, wrap examples in tags, and tell the model what to do instead of what not to do. Every prompt below uses three parts: a rule, an output format, and your text inside tags.
The prompts
Replace TEXT with your passage. Keep the tags.
1. Minimal proofread with a change list
Proofread the text in <text> tags. Fix only grammar, spelling, punctuation and capitalization errors. Keep every other word, the order of sentences and my tone exactly as written. Return two parts: (1) the corrected text, (2) a numbered list of every change as "original -> corrected (error type)". If you find no errors, say so.
<text>TEXT</text>
2. Category audit
Check the text in <text> tags for these error types, one at a time: homophone mix-ups, missing or repeated words, sentence fragments, run-on sentences, comma splices, subject-verb agreement, mixed constructions, parallelism in lists, unclear pronoun reference, and apostrophe misuse. For each type, quote the problem sentence or write "none found". Do not rewrite anything yet.
<text>TEXT</text>
3. Flag only, no edits
List every sentence in <text> tags that may contain a grammar error. For each, quote it, name the suspected error, and explain in one sentence. Do not change my text. I will decide what to fix.
<text>TEXT</text>
4. Explain the rule
Correct the grammar in <text> tags. For each correction, give the rule in plain language and one short example sentence of the rule used correctly. I am learning, so explain rather than just fix.
<text>TEXT</text>
5. One error type only
In <text> tags, correct only comma splices and run-on sentences. Leave all other issues untouched, even if you notice them. Show the changed sentences before and after.
<text>TEXT</text>
6. Ask before guessing
Proofread <text> tags. If a correction depends on what I meant, do not guess: list it under "Questions for the author" with the two possible readings. Apply only the corrections that have one clear answer.
<text>TEXT</text>
7. Formal email pass
Check this email draft for grammar and punctuation only. Keep it concise and keep the greeting and sign-off as written. Return the corrected email, then up to five bullet points of the most important fixes.
<text>TEXT</text>
8. Non-native writer pass
I am writing in English as an additional language. Correct grammar, article use (a, an, the) and preposition errors in <text> tags. Keep my vocabulary unless a word is wrong. Show each change in a table with columns: original, corrected, reason.
<text>TEXT</text>
A worked example you can check
Here is a test passage with six seeded problems:
Their going to announce the results on friday. The applications for the grant was reviewed in march, it took longer than expected. Each applicant must submit there portfolio, a cover letter, and to provide two references.
The correct result, which a good response to Prompt 1 should match:
They're going to announce the results on Friday. The applications for the grant were reviewed in March; it took longer than expected. Each applicant must submit their portfolio, a cover letter, and two references.
| Original | Corrected | Error type |
|---|---|---|
| Their going | They're going | Homophone (contraction) |
| friday, march | Friday, March | Capitalization |
| applications ... was | applications ... were | Subject-verb agreement |
| march, it took | March; it took | Comma splice |
| there portfolio | their portfolio | Homophone (possessive) |
| and to provide two references | and two references | Parallelism in a list |
Compared word by word, the two passages differ by exactly these edits and no others, and the word count falls from 36 to 34. If a model's answer changes anything else, such as swapping "expected" for another word, it broke Prompt 1's rule.
Checking the result yourself
Purdue OWL advises reading your text aloud, and says that any time you stumble or have to reread, you should revise that section. Do this with the corrected version. Also keep a list of your own frequent errors, as Purdue suggests, and paste it into Prompt 2 in place of the default categories so the check targets your habits.
Two quick checks catch most over-editing: compare word counts (a large change means something was rewritten), and confirm that every difference you can see appears in the change list. Anything that does not should be rejected.
Take this further with Oakgen Agent
Proofread pasted text or an uploaded DOCX or PDF (up to 100 pages), list each change with its error type, and produce a corrected document file.
Looking for a recap of one specific YouTube video? The YouTube Summarizer does that today.
Open this task in Oakgen Agent
The prompt is filled in for you. Nothing is sent until you press send.
Frequently asked questions
How do I stop ChatGPT from rewriting my whole text during a grammar check?
State the limit as a rule and say what to do instead. Anthropic's prompting guide recommends telling the model what to do rather than what not to do, so write 'Change only words that are grammatically wrong; keep every other word exactly as written' and ask for a change list. Then compare the output with your original. If you see edits that are not in the change list, ask for a corrected version again. These prompts follow vendor guidance, and the page has not measured how often any model breaks the rule.
Which grammar errors should I ask it to look for?
Purdue OWL's proofreading guide covers a practical list: homophones such as your and you're, left-out or repeated words, sentence fragments, run-on sentences, comma splices, subject-verb agreement, mixed constructions, parallelism, pronoun reference and apostrophe use. Prompt 2 below names these categories so the model checks each one. Keeping your own list of frequent mistakes, as Purdue suggests, lets you tailor it.
Is it safe to paste confidential text into a chatbot?
That depends on the service's data terms and your organisation's rules, neither of which this page assesses. Remove names, figures and anything you are not allowed to share before pasting, and check the privacy settings of the product you use.
Can ChatGPT replace a human proofreader?
This page does not claim that. Purdue OWL advises reading your text aloud slowly and tracking your own frequent errors, which are habits no prompt replaces. Use the prompts as a second pass, verify each change against your intended meaning, and have a person review anything high-stakes such as legal, medical or published text.


