ComicsBrowseMy shelfCreate
In progress|
Writing with AI

How to Revise Your Novel with AI: Reread Your Own Draft in the Order Story → Scene → Sentence → Surface, with Prompts and Limits on How Much the AI Rewrites

12 min readOriginal Novel Editorial Team

To revise your own novel with AI, reread it four times, moving from rough to fine, and ask the AI only for work that measures the manuscript: summarizing, sorting, counting and searching. That way you polish the draft and keep your own style. Using a short scene on a station platform, this article shows a prompt for each pass and how to use only what the AI points out, without taking its rewrites.

At a jeweler's workbench, a book character polishes a lumpy stone made of crumpled manuscript paper with a cloth, slowly turning it into a shining gem, while a small lamp with a magnifier lights up the work

Key takeaways

  • Revise in four passes: story → scene → sentence → surface. Big fixes wipe out small ones, so work from the outside in
  • Give the AI the jobs that measure: summarizing, sorting, counting and searching. Make the fixes with your own hands
  • In pass 1, have the AI summarize without telling it your aim, then compare what came across with your aim yourself
  • Don't paste in the AI's rewordings; use only the substance of what it points out. Save versions and compare before and after

How to revise your novel with AI: four passes, from rough to fine

If you use AI to revise your own novel, the shortcut is this: don't try to fix everything in one go. Reread the draft four times, in the order story → scene → sentence → surface, and give the AI just one job each time. What you hand to the AI is work that measures the manuscript: summarizing, sorting, counting and searching. Which sentences stay, and how they get reworded, is decided by the person who wrote them.

The order matters because big fixes wipe out small ones. Say you polish sentence ③ of the example below, “The train still hadn't come.” If the second pass then shows that it repeats the same situation, the whole sentence gets cut anyway.

Pass 1: Story (the scene's role) Pass 2: Scene (roles and order) Pass 3: Repeats and rhythm Pass 4: Surface Typos, spelling variants, clashes with earlier chapters The job you give the AI Pass 1: summarize the scene Pass 2: sort the sentences by role Pass 3: count tenses and repeated words Pass 4: find variants and clashes You make every fix yourself
Figure 1. The farther out the frame, the bigger the fix. Rereading from the outside in means you don't end up throwing away a sentence you've already polished.

Revise, edit, proofread, copyedit: the word you use changes what the AI gives back

Asking an AI to “revise” your writing and asking it to “edit” or “correct” it gets you different answers.

The Japanese word for revising, suikō (literally “push-knock”), comes from a story about the Tang-dynasty Chinese poet Jia Dao. He couldn't decide whether his line “the monk pushes the gate under the moon” should say “pushes” or “knocks,” asked the scholar Han Yu, and settled on “knocks.” Tensaku, by contrast, is defined by the Japanese dictionary Digital Daijisen as reworking someone else's poems, writing or exam answers by adding or cutting words (the Nihon Kokugo Daijiten does not limit it to “someone else's”).

Revising, editing, proofreading and copyediting (based on how Japanese dictionaries explain the Japanese words, with ways to ask the AI)
WordJapanese dictionary explanation (summary)How you might phrase it to the AI
Revise (suikō)To examine the wording of a poem or text carefully and rework it“Summarize this scene,” “Sort these sentences by role” (the AI measures; you make the changes)
Edit / correct (tensaku)To rework someone else's writing by adding or cutting words (Digital Daijisen)“Edit this” (tends to bring back rewritten sentences)
Proofread (kōsei)To compare words and text and fix errors; in printing, to check proofs against the manuscript“List the typos and inconsistent spellings”
Copyedit (kōetsu)To examine a manuscript for errors and gaps, consider them and correct them“Find anything that contradicts the details set up in earlier chapters”

Asking an AI to read your work and give you feedback (editing and critique) is covered in ChatGPT Prompts for Editing a Novel and Honest Feedback.

The example we'll follow: waiting for her brother on a station platform

We'll follow one short scene all the way through the four passes. The writer's aim: “Her brother was always late for everything, but today he got there first. His younger sister Nana is thrown by how he has changed.” The numbers at the start of each sentence are only there for the explanation.

① The station platform was very cold. ② Nana regretted forgetting her gloves. ③ The train still hadn't come. ④ Her brother, whom she was meeting for the first time in three years, had always been late for everything. ⑤ Nana looked at her smartphone. The train still hadn't come. ⑥ Just then, a voice behind her called, “Nana.” ⑦ She turned around and saw her brother standing there, and Nana was very surprised.

There are no mistakes, and it makes sense. But whether the aim comes across is hard for the writer to see.

Pass 1, the story: have the AI summarize, then compare what came across with your aim

First, check whether the scene does its job in the story. Ask the AI for a summary, so it puts into words what a reader actually takes away.

The key is not telling the AI your aim first. AI answers can get pulled toward what the person asking already thinks. Rather than explaining your aim and then asking “Does this come across?”, get the summary first and compare it with your aim yourself. That's more reliable.

Read the scene below and answer the questions. I don't need any suggested changes.

1. What happens in this scene, in 3 lines
2. Two new things the reader learns about the characters by the end
3. One thing the reader wants to know next at the end of this scene

[Summary of the story so far] (3–5 lines)
[This scene]
(paste here)
Figure 2. A pass 1 prompt. Don't state your aim; ask only what came across.

Ask for a summary of the example and you get something like: “Nana is waiting for her brother on a cold platform. He shows up, and she is surprised.” What's missing is that her brother got there first. The voice in ⑥ alone doesn't tell us how long he had been there. So the writer decided to show, in the text itself, that he had been waiting on a bench. The writer made that call; the AI only reported what came across.

Pass 2, the scene: have the AI sort sentences by role and find what isn't needed

Next comes the order within the scene. When the AI sorts the sentences by role, you start to see sentences doing the same job twice and sentences that don't connect to anything.

Go through the scene below from the top, one sentence at a time, and label each sentence by its role. The role is one of: setting / feeling / past / event.

How to answer:
- Numbered: “sentence number / role / which later sentence it connects to (‘none’ if none)”
- At the end, list any places where sentences with the same role come one after another
- Don't rewrite any sentences

[Scene]
(paste here)
Figure 3. A pass 2 prompt. The AI doesn't rewrite anything; it only returns roles and connections.

For the example, the sorting showed three things:

  • ③ and the second half of ⑤ both repeat the same situation: the train still hasn't come
  • The gloves in ② don't connect to any later sentence
  • “Very surprised” in ⑦ simply states the feeling outright instead of showing it

One of the repeats gets cut. The gloves stay: the brother will say, “Forgot your gloves again, huh?” That one line shows both that he remembers his sister's habit and that he got there first and waited.

From pass 2 on, one scene at a time is a comfortable amount to give the AI. Across the Japanese chapters of the stories published on Original Novel (227 chapters, counted on November 14, 2026), the median chapter length was 3,806 Japanese characters. A chapter holds several scenes, so giving the AI one scene at a time rather than a whole chapter tends to get you a more detailed sort.

Pass 3, the sentences: have the AI count tenses and repeated words, and do the rewording yourself

Pass 3 looks at each sentence. What the AI is good for here is counting: the kind of numbers human eyes easily miss.

Count the following in the scene below and put the results in a table. Don't suggest any changes.

1. The tense of each sentence (past, present, etc.) and how many sentences in a row use the same tense
2. Words that appear 2 or more times (“very,” “Nana,” etc.) and how many times each
3. The longest sentence and the shortest sentence

[Scene]
(paste here)
Figure 4. A pass 3 prompt. The only job you give the AI is counting.

For the example, the result was: all eight sentences (⑤ is two sentences) are in the past tense, “very” appears twice, and “Nana” is the subject three times. If you then try asking the AI to reword ① as well, you get a neat suggestion like the one below.

Three versions of sentence ①. The AI's suggestion wasn't used; only its observation was.
VersionSentenceWhat we noticed
First draftThe station platform was very cold.It only says “cold” with “very.” Past tense
The AI's reworded versionThe frozen night air filled the platform, mercilessly stealing the warmth from Nana's fingertips.Polished, but heavy, and it doesn't match a writer who builds with short sentences
The writer's own revisionThe wind on the platform chills her from the fingertips up.Shows the cold through the body and switches from past to present tense. It also leads into the gloves sentence that follows

If you take the AI's version, that one sentence ends up in someone else's voice. What you use is the observation, “it only says ‘cold’ with ‘very,’” and you find the words yourself.

Pass 4, the surface: have the AI find variants and contradictions with earlier chapters

Last, check for inconsistent spellings and word choices (“phone” in one place, “smartphone” in another, for example) and for anything that contradicts earlier chapters. To find contradictions, give the AI a reference sheet with names, what characters call each other, ages, dates and so on.

Check the scene below against the reference sheet, one line at a time. Don't change the text.

How to answer:
- For each line of the reference sheet, a table row: “reference sheet line / how this scene puts it (‘none’ if it doesn't come up) / matches, contradicts, or please check”
- After the table, list any words where the same thing is written two different ways (e.g. smartphone and phone)

[Reference sheet] (bullet points: names, what characters call each other, ages, dates, belongings, etc.)
[Scene]
(paste here)
Figure 5. A pass 4 prompt. Hand it over together with your reference sheet.

In the example, the earlier chapters said “phone,” but this scene said “smartphone.” The AI also asked whether “for the first time in three years” was right, since the reference sheet says “the brother left home four years ago, in spring.” Checking the sheet again, the writer found that Nana had seen him once, a year after he left, so it was correct. Take what the AI flags as “things to check,” and find the answers in your own notes.

Original Novel's story-making works the same way: each story keeps a blueprint, its foreshadowing and a record of established facts, and these are handed to the AI every time a chapter is written, doing the same job as the reference sheet in pass 4. On top of that, before each chapter is published, AI checks it for contradictions with earlier chapters and for mistakes in names and numbers, and the site's operators review any problems it flags.

After all four passes, the example turned into this:

The wind on the platform chills her from the fingertips up. She left her gloves at home again. Her brother, whom she is meeting for the first time in three years, has always been late for everything. She checks the time of the next train one more time.
“Nana.”
The voice comes from the bench behind her. Her brother is sitting there, holding two cans of hot cocoa.
“Forgot your gloves again, huh?”

The word “surprised” is gone, yet the fact that her brother got there first, and that he remembers her habit, comes across more clearly than in the first draft. Every change was chosen by the writer, based on facts the AI pointed out.

Where to draw the line so the AI doesn't rewrite too much

Draw these lines before you start, and you won't waver:

  • Ask the AI to rewrite only when you're stuck for wording: otherwise, keep it to the four jobs of summarizing, sorting, counting and searching
  • Don't paste suggestions in as they are: take the substance of the observation and choose the words again yourself
  • Finish by reading aloud: your own ear is a better judge than the AI of whether a revised sentence sticks out

Tips on how to ask, such as marking the sentences you don't want changed ahead of time or narrowing what you want checked to one thing, are collected in ChatGPT Prompts for Editing a Novel and Honest Feedback.

A book character reads a revised manuscript aloud, picking up just one of the suggestion cards laid out on the desk and pushing the rest aside
Figure 6. Lay the suggestions out and look them over, but you do the choosing. Finish by checking it aloud.

Alongside these lines, keep up one more habit: put the before and after side by side and look only at what changed. Over several rounds of back-and-forth, words can change even in places you didn't ask about. Save a version every time you revise (name it, e.g. “first draft,” “after pass 1”) and compare it with the previous one.

In Google Docs, Tools > Compare documents lets you compare two documents. The differences appear as suggested edits in a third document, and you can accept or reject each one. Without a tool like that, ask the AI to “list only the sentences that changed between the previous version and this one,” then write down the reason for each change yourself. A change you can't give a reason for is a candidate to undo.

Before you paste in an unpublished manuscript, also check how the text you enter will be handled. In the Gemini app, for example, Google says that when the History setting is on, your conversations may be used to improve its services, including training generative AI models. How the settings differ across the three companies is summarized in a table in ChatGPT Prompts for Editing a Novel and Honest Feedback.

When to stop revising

With AI, there's no end to suggested changes, so it's hard to see when you're done. A good sign is when pass 4 only turns up small surface notes, and asking for the pass 1 summary again gives you back the same thing as your aim. Beyond that, more rewording mostly just makes a different sentence, not a better one.

The pass 1 summary prompt works on other people's stories too. Paste in one chapter of a published story, have the AI summarize it, and then look in the text for the characters' feelings that appear in the summary. If a feeling isn't put into words anywhere, that's a place where the writer shows instead of explains. It's a model to follow when you cut the “very surprised” from your own manuscript.

Stories on this site to read with a reviser's eye

See more: Stories on this site to read with a reviser's eye ›

These are stories published on this site. As you read, look for places where feelings are shown rather than put into words, and for the last line of each scene.

Sources (date checked)

  1. Suikō – to polish and revise (dictionary entry)Kotobank (Digital Daijisen, Nihon Kokugo Daijiten) / Japanese / Checked Nov 18, 2026
  2. Tensaku – to correct and edit (dictionary entry)Kotobank (Digital Daijisen, Nihon Kokugo Daijiten) / Japanese / Checked Nov 18, 2026
  3. Kōsei – proofreading (dictionary entry)Kotobank (Digital Daijisen, Nihon Kokugo Daijiten) / Japanese / Checked Nov 18, 2026
  4. Kōetsu – copyediting (dictionary entry)Kotobank (Digital Daijisen, Nihon Kokugo Daijiten) / Japanese / Checked Nov 18, 2026
  5. Towards Understanding Sycophancy in Language ModelsarXiv / Checked Nov 18, 2026
  6. Compare documents in Google DocsGoogle Docs Editors Help / Japanese / Checked Nov 18, 2026
  7. Gemini Apps Privacy HubGemini Apps Help (Google) / Japanese / Checked Nov 18, 2026

FAQ

Can I submit a novel I revised with AI to a writing contest or a story-posting site?

How AI-assisted work is treated differs from contest to contest and site to site. Before you submit, check the entry rules or terms of use for what they say about AI. Whether those rules cover using AI only as a revision tool is also something to judge from the explanation given by the place you're submitting to.

Can I use these revision prompts for essays or reports, not just fiction?

Yes. Pass 1 (summarize and compare with your aim), pass 3 (counting) and pass 4 (finding inconsistencies) work as they are for any kind of writing. Only for pass 2, swap the roles “setting / feeling / past / event” for ones that fit the piece, such as “claim / reason / example / conclusion.”

Should I revise a draft right after I finish writing it?

Right after you finish, you tend to read what you intended into the sentences, so it's best to let some time pass before rereading. If you can't wait, asking the AI for the pass 1 summary lets you check, from the outside, only what is actually written.

How should I go about revising a whole long novel?

For pass 1 only, give the AI the chapter summaries in order and ask, “If this chapter were cut, what problems would that cause in later chapters?” A chapter whose answer is “nothing in particular” is a candidate to shorten or merge with another. From pass 2 on, give it one scene at a time, and keep adding to the reference sheet you use in pass 4 each time you revise a chapter.

#prompts#ChatGPT#Claude

Read next