Using AI to Edit Your Own Writing Without Losing Your Voice

Why naive "edit this" prompts flatten your voice

If you paste a page into an LLM and say "edit this" or "make this better", you’re giving it a blank cheque. The safest thing it can do is converge toward generic, risk-free prose.

Models like Claude Opus 4.7 or GPT-5.1 have been trained on oceans of writing. When you don’t specify what to protect, they optimise for average readability and politeness. Individual quirks, your rhythm, strange metaphors, side comments, look like noise to be cleaned up.

You also create an incentive problem. With a vague prompt, the easiest way for the model to prove it “worked” is to change a lot. Big rewrites look impressive, even if your original was 80% fine.

So the default behaviour is:

You can’t fix that with a magic better-prompt. You fix it by changing the contract: narrow the task, fix the mode, and make the model suggest instead of overwrite.

The three editing modes: line, structure, fact

When you say "edit", you might mean three different jobs. Models are bad at guessing which one you care about in a given moment.

Line-level

Sentence-by-sentence tuning for clarity, rhythm, and word choice. Same ideas, same order, micro-level changes only.

Structural

Reordering or regrouping sections, sharpening the lede, changing what you emphasise. Paragraphs move; content may be cut or expanded.

Fact-check

Checking whether claims, dates, numbers, quotes, or technical statements are plausible and sourced. No stylistic changes.

If you bundle all three into one request, you get an AI that:

…all at once, with no visibility into which change served which purpose.

Treat AI like a set of small, specialised tools instead of a single "make it better" button. You will move slower per pass but much faster per improvement. Each run should answer one kind of question: do these words scan better, is this structure clearer, or is this claim actually true? If you mix those up, you can’t tell whether to accept, reject, or debug a change, and you lose your voice as collateral damage.

We’ll stay in one mode at a time and tell the model which one we’re in.

Find your starting mode: where are you actually stuck?

Before you open an AI tab, decide what kind of stuck you are. This matters more than which model you choose.

You’re probably in one of three situations:

For a first run, most writers in your position should start with line-level. You already have a voice and a complete draft; you just want sharper sentences.

We’ll build a small loop around a single paragraph. Once that feels reliable, you can scale the same pattern to the rest of your work.

First attempt: a line-level edit loop on one paragraph

Pick a paragraph you wrote recently, 80 to 150 words is ideal. Paste only that paragraph into the model.

Here’s a starter prompt you can adapt:

You are an editing assistant. Stay strictly in line-editing mode.

Rules:
- Do NOT rewrite the whole paragraph.
- Do NOT change my tone, point of view, or content.
- Keep my metaphors, examples, and asides.

Task:
1. For each sentence, give up to 3 alternative phrasings that improve clarity and rhythm.
2. Number the sentences and the options like 1a, 1b, 1c, 2a, etc.
3. For each option, add a 1-sentence explanation of what you changed and why.

Return ONLY suggestions, not a full rewritten paragraph.

Here is the paragraph:
---
[PASTE YOUR PARAGRAPH]
---

This does three important things:

1. 1
   Locks the mode: no structure, no facts, just sentences.
2. 2
   Forces suggestions: "up to 3 options" per line, not a one-shot answer.
3. 3
   Requires explanations so you can judge the reasoning, not just the phrasing.

Run this once before you read on so you have real output in front of you.

Reading the feedback: what to accept, what to ignore

Look at the model’s suggestions as proposals, not solutions. Your job is to be the copy chief.

Good signals:

Bad signals:

On this first attempt, pick at most three suggestions to accept. Manually paste them into your paragraph and read the whole thing aloud. If a change sounds like you on a good day, keep it. If you’d never say that phrase out loud, roll it back.

Structural passes without losing your tone

Once line-level edits feel safe, you can have the model help with structure. The trick is to separate analysis from rewrite.

Start with this kind of request on a full section or page:

Read this section and stay in structural-editing mode.

Rules:
- Do NOT rewrite sentences.
- Do NOT change my voice.

Task:
1. Outline the current structure in 5-7 bullet points.
2. Point out 2-3 places where reordering or trimming could improve focus.
3. Suggest 2 alternate structures as outlines only.

Text:
---
[PASTE SECTION]
---

Your first pass is just reading the model’s outline. Does it match what you thought you wrote? If its outline misrepresents your argument, your draft is ambiguous; that’s useful feedback.

Only after you agree on an outline do you ask for actual rewrites:

Using structural option 2, propose a revised order of my existing paragraphs.

- Do NOT add new paragraphs.
- Do NOT change sentence wording.

Return a numbered list indicating the new order, with 1-sentence notes on why each move helps.

You then apply this reordering yourself in your editor. That manual step helps keep your tone intact.

Practical fact-checking with a model that still hallucinates

Modern LLMs are still not reliable sources of truth, even when they sound certain. For factual editing, think of them as research assistants, not authorities.

A simple protocol:

1. Mark the claims in your draft that actually matter: numbers, dates, named studies, strong assertions.
2. Ask the model to list these claims and say which ones look risky or need sources:

Highlight every factual claim in this excerpt.
For each, tell me:
- why it might be wrong or outdated
- what kind of source could confirm it (official stats, primary research, documentation, etc.)

Do NOT rewrite any prose.

1. For any non-trivial claim, ask the model where you’d verify it. Use it to generate search queries and source types, not final answers.
2. Go to primary sources yourself: official docs, papers, reputable databases.

If the model does propose corrections, treat them as hypotheses. Ask, "Where would I check this?" not "Do I believe this?"

Worked example: a real paragraph before and after

Let’s walk through a full mini-loop on a deliberately imperfect paragraph.

Original paragraph

When you throw your draft into an AI and beg it to just "fix" things, what usually comes back isn’t really your writing anymore, it’s like a smoothie made from ten thousand blog posts, technically drinkable but kind of depressing. The sharp edges you liked are filed off, the weird little turns of phrase are gone, and somehow everything sounds like whoever writes airport business books.

We run a slightly shortened line-edit prompt on just this paragraph.

AI suggestions for the first sentence (imagined but realistic):

1a. "If you toss your draft into an AI and beg it to 'fix' things, what usually comes back isn’t really your writing anymore; it’s a smoothie made from ten thousand blog posts, technically drinkable, but kind of depressing."

1b. "When you hand your draft to an AI and ask it to simply 'fix' things, what you get back rarely feels like your writing; it’s more like a smoothie of ten thousand blog posts, drinkable but quietly depressing."

1c. "Drop your draft into an AI and ask it to 'fix' things, and it returns something that barely feels like yours: a smoothie of ten thousand blog posts, technically drinkable but oddly depressing."

Assume the model explains that it fixed a comma splice and tightened phrasing. Suppose we accept 1a because it keeps the metaphor and cadence but repairs the run-on.

Now for the second sentence, we ask the model to focus on rhythm, again in options. We end up accepting a version like:

The sharp edges you liked are filed off, the weird little turns of phrase disappear, and somehow everything ends up sounding like an airport business book.

Resulting paragraph (after manual selection):

If you toss your draft into an AI and beg it to "fix" things, what usually comes back isn’t really your writing anymore; it’s a smoothie made from ten thousand blog posts, technically drinkable, but kind of depressing. The sharp edges you liked are filed off, the weird little turns of phrase disappear, and somehow everything ends up sounding like an airport business book.

Notice what did not change:

That’s the goal: you use the model as a precise instrument, not a co-author.

Retrying smarter when your first attempt goes wrong

If your first pass felt off, too bland, too formal, or too heavy-handed, treat that as a prompt bug, not a personal failure.

Here are three common failure modes and how to retry:

1. The model rewrote the whole thing anyway.

Tighten the guard rails. Add lines like:

If you cannot follow these instructions exactly, stop and ask clarification questions.
Do NOT output a full rewritten paragraph under any circumstances.

Also reduce the scope further: work on two sentences, not a full paragraph.

2. Everything sounds like corporate copy.

Name your style explicitly and forbid formalising:

My voice is informal, first-person, with occasional dry humour.
Do NOT make the tone more formal or more enthusiastic.
When in doubt, keep my original wording.

Ask for "minimal edits" or "edits that change at most 5-7 words per sentence".

3. Suggestions are timid and barely different.

In that case, you can safely ask for bolder options:

Give me 2 safe options and 1 bold option per sentence.
The bold option can change structure more, but must keep my metaphors and stance.
Label them clearly as SAFE or BOLD.

Over a few sessions, you’ll end up with a short, personal editing prompt that behaves the way you expect.

Cheatsheet: three modes, one workflow

Here’s a compact reference tying the three modes together.

Keep this table next to your editor the next few times you run AI-assisted edits. If a prompt output doesn’t match the "What the AI returns" column, your instructions were too loose.