Before you start
You need a draft and a chat app. Before the first request, copy the draft into a second file. Every step below assumes you can go back to your own words.
Editing covers four jobs: fixing spelling and grammar, making sentences clearer, cutting length and changing tone. Ask for one job per request. A request that mixes all four gives you a text you cannot compare with the original.
An edit can change your wording even when you did not ask for that. In a paper by Geng and Trotta on Hugging Face Papers, GPT-3.5 was asked to "Revise the following sentences:" in 10,000 arXiv abstracts from April to July 2022. The words "is" and "are" appeared less often afterwards: in computer science, "is" fell from 2.01 to 1.73 and "are" from 1.00 to 0.83 per abstract. The authors note that other models may have different word preferences. Wikipedia's guide to AI writing cites the same paper.
Ask what is wrong before you ask for a rewrite
A rewrite hides its reasons, while a diagnosis shows them and leaves the choice to you. Anthropic's doc-coauthoring skill re-reads a nearly finished document and checks four things. The prompt below turns those four checks into a request.
Read the draft below as an editor. Do not rewrite anything yet.
Report what you find under these four checks, quoting the sentence each time:
1. Flow and consistency across sections
2. Redundancy or contradictions
3. Anything that feels like "slop" or generic filler
4. Whether every sentence carries weight
Leave my facts, quotes and numbers alone.
[paste your draft]- Read the list and strike every finding you disagree with.
- Pick the one or two problems that matter most for your reader.
- Ask for a fix to those only, using the next section's kind of request.
Ask for one specific change at a time
The skill gives two examples of the feedback it wants. One removes a named bullet because another bullet already covers it, and the other is "Make the third paragraph more concise". Each names a place and a change.
Its rule for edits is to change the named text and "never reprint the whole doc". A chat app has no such rule, so add it to your request: return only the paragraph you named, and leave the rest word for word.
| Vague request | Specific request | Why it works |
|---|---|---|
| Make it better. | Make the third paragraph more concise. | It names the place and the change. This one is the skill's own example. |
| Fix the intro. | Cut the first two sentences and open on the date. | It says what to remove and what comes first. |
| Make it sound professional. | Rewrite this for a finance manager in plain words, and keep every number. | It names the reader and what must not move. |
| Improve the flow. | Move the pricing paragraph before the setup paragraph and change nothing else. | It asks for one move, with a limit. |
Get edits as suggestions you can accept or reject
A rewritten paragraph hides what changed, and a change you cannot see is one you cannot refuse. Ask for the edits as a list instead.
Word processors show the same idea as tracked changes. Anthropic's docx skill tells an agent how to edit Word files. When it redlines, it wraps each change in tracked-change markup, and its validator, run with the author's name, reports any text changed without that markup. The file calls that slip "easy to do by accident and invisible in the accepted view".
Suggest edits to the text below as a numbered list. For each edit give:
Original: the sentence as I wrote it
Revised: your version
Why: one short reason
Do not change anything you did not list. I will accept or reject each edit myself.
[paste your draft]- Paste your draft with the prompt. The app returns a numbered list.
- Go down the list and mark each edit accept, reject or rewrite in your own words.
- Apply the accepted edits in your own file, then read the whole text again, because edits that are fine alone can clash.
Shorten it, or change the tone
Give length as a number and the reader as a person. OpenAI's guide says a senior exec, legal, a developer, a customer and an internal team "each implies different language and detail level", so put the reader in the request. To cut, ask what can go. The skill's check is to ask whether anything can be removed without losing important information, and to ask it after three rounds in a row with no substantial change.
Do not count on a tone setting. OpenAI's cookbook page on prompt personalities is written for developers who set an agent's personality in its prompt. It says a personality does not override a task's output format: if you ask for an email, the model should follow your instructions and the task context rather than the selected personality. The page says nothing about the tone setting inside ChatGPT, so state the tone in the request itself.
| Job | Say |
|---|---|
| Shorten | Cut this to 80 words. Keep the deadline and the request. List what you removed. |
| Cut a section | Can anything in section 2 be removed without losing important information? Show me the text, and do not delete it yet. |
| Warmer | Rewrite for a customer who is upset. Keep the facts and the length, and open with one sentence that acknowledges the delay. |
| Firmer | Rewrite for a supplier who missed two deadlines. Plain words, no apology, and keep the new date. |
| Plainer | Rewrite for a reader who is new to the topic. Replace jargon with everyday words and keep every number. |
Check the edit against your original
An edit can change a fact while it fixes a sentence. Paste your original and the edited text, and run the check from OpenAI's guide. The five lines under the first one are the guide's own, and the first line is added for comparison.
Compare the edited text with my original, then check the edited text for:
- Accuracy: Does it avoid making up facts?
- Completeness: Did it include the required points?
- Format: Does it match the requested structure?
- Tone: Is it appropriate for the audience?
- Assumptions: Are any guesses clearly labeled?
[paste your original, then the edited text]- Read the edited text aloud. The humanizer skill puts that in its step for checking a draft.
- Put back any phrase of yours that the edit replaced and that you want to keep.
- If voice matters, paste your sample again. The humanizer skill matches its sentence length, word choice, punctuation, openings and transitions.
See an edit done inside a document
Anthropic's Claude channel on YouTube has a video titled "Claude for Word: Turn a draft into a finished document". The title points to Claude working inside Word, which this lesson does not need. Every prompt above works in a plain chat window.
Common mistakes
- Saying "make it better". The app has to guess what better means, and its guess may not be yours.
- Pasting your only copy and accepting the result. Keep the original in a second file.
- Letting the app touch facts. Run the five-point check and compare each figure with your notes.
- Asking for tone, length, cuts and grammar in one request. You cannot tell which change helped.
- Accepting every suggestion. Reject any edit you cannot explain, and keep your own sentence when the two say the same thing.
Try it yourself
- Take a draft you wrote this month, copy it to a second file and run the four-check prompt.
- Ask for the original, revised and why list, and accept no more than half of it.
- Ask for a version 20 percent shorter, with a list of what it removed.
- Run the five-point check against your original, then read the result aloud.
Next: signs of AI writing
Edited text can still read as machine-made. Signs of AI writing lists what readers notice, with a plain fix for each sign.
If you have not started from your own notes yet, drafting with AI comes first in this course.
Check yourself
What should you do before the first edit request?
Show answer
B. Copy the draft to a second file and name what the app may not change Every later step assumes you can go back to your own words, and naming what must stay stops the app from touching facts and quotes.
Which request names both a place and a change?
Show answer
C. Make the third paragraph more concise It is the doc-coauthoring skill's own example of useful feedback, and it points to one paragraph and one change.
Why ask for edits as original, revision and reason?
Show answer
D. You can accept or reject each change instead of receiving one rewritten text A rewritten paragraph hides what changed, while a list lets you decide on each edit.
An agent's prompt sets a personality, and you ask it for an email. What does OpenAI's cookbook page say the model should do?
Show answer
A. Follow your instructions and the task context rather than the personality The page says a personality does not override a task's output format. It is about agent prompts, not ChatGPT's tone setting, so tone belongs in the request itself.
Frequently asked questions
How do I edit my writing with ChatGPT without changing my voice?
Name what must stay, ask for one change at a time, and ask for edits as a list of original, revision and reason, so you accept or reject each. Paste a sample of your writing if you want the edit to match it.
Can AI proofread without rewriting my sentences?
Yes, if you limit the job. Ask it to fix spelling, grammar and punctuation only, to leave the wording alone and to list each fix. Then check the list against your original.
How do I see what the AI changed?
Ask for the changes as a numbered list with the original, the revision and a reason. In a Word file, an agent can mark changes as tracked changes, and when it redlines, Anthropic's docx skill has a check for text changed without that markup.
What is the best prompt for editing an essay?
OpenAI's guide says there is no single perfect prompt template. A scoped one works: say who reads the essay, name one job and name what to keep. The four-check prompt in this lesson is a good first pass.
Why does my text sound like AI after I edit it with AI?
An edit can swap your wording for the app's. In one paper, GPT-3.5, asked to revise sentences, used "is" and "are" less often. The next lesson lists the other signs and how to fix each.
Can AI edit without making up facts?
Not reliably. Tell it to leave facts alone, then check every name, date and figure against your original, and run the five-point check from OpenAI's guide.
Sources
- Anthropic skills on GitHub: doc-coauthoring https://github.com/anthropics/skills/blob/main/skills/doc-coauthoring/SKILL.md
- OpenAI Cookbook: ChatGPT Enterprise Prompting Guide https://developers.openai.com/cookbook/examples/chatgpt/chatgpt_prompt_guide/chatgpt_prompt_guide
- OpenAI Cookbook: Prompt Personalities https://developers.openai.com/cookbook/examples/gpt-5/prompt_personalities
- Anthropic skills on GitHub: docx https://github.com/anthropics/skills/blob/main/skills/docx/SKILL.md
- GitHub: the humanizer skill (blader/humanizer) https://github.com/blader/humanizer/blob/main/SKILL.md
- Hugging Face Papers: Is ChatGPT Transforming Academics' Writing Style? (Geng and Trotta) https://huggingface.co/papers/2404.08627
- Wikipedia: Signs of AI writing https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
- Claude on YouTube: Claude for Word, turn a draft into a finished document https://www.youtube.com/watch?v=x80HVKbZrno