Using ChatGPT to Edit Your Novel: Prompts and Limits

Yes, you can use ChatGPT to edit your novel, and it is genuinely useful at scene level. Paste in a chapter and it will give you line notes, flag a confusing paragraph, brainstorm fixes for a sagging middle, or answer a pointed question about whether a character's motive holds up.

Where it struggles is the whole book. A general chatbot tends to lose track across chapters, praise more than it criticises, quietly rewrite your voice when you asked for feedback, and give different answers each time you run the same request. The fix is to ask for diagnosis rather than rewrites, and to feed it the book in a structured way.

StoryWith AI is our own product, so weigh what we say about it accordingly.

Key takeaways

• General chatbots are good at one scene or chapter at a time: line notes, brainstorming, and consistency questions about text you paste in.
• Whether a full 90,000-word novel even fits depends on the plan and mode. On ChatGPT Plus, the fast mode's window is far too small; the reasoning mode's is big enough on paper, close to OpenAI's stated input maximum.
• Fitting is not the same as reading well. Long inputs still lead to missed details, especially in the middle of the book.
• Ask for diagnosis, not rewrites. Prompts that forbid rewriting keep your voice intact.
• For a whole-book pass, build chapter summaries and a running story bible, then question those.
• Check training settings before pasting unpublished work. On ChatGPT and Gemini, switch training off yourself.

Can ChatGPT actually edit a novel?

It can do some editing jobs well and others poorly. Think of it as a sharp, fast reader who has only ever seen the pages you just handed over.

Jobs it handles well on a pasted scene or chapter:

• Line-level notes. Repeated words, tangled sentences, filter words, dialogue tags that pile up.
• Clarity checks. Asking where a reader might lose track of who is speaking or where the scene is set.
• Brainstorming. Five ways to raise the stakes in chapter 12, or three options for a stronger chapter ending.
• Targeted consistency questions. "In this chapter, does Mara's account of the fire match what she said in the excerpt below?"

Jobs it handles badly without help:

• Tracking a subplot across 30 chapters.
• Noticing that a gun introduced in chapter 4 never fires.
• Telling you, consistently, which five problems matter most in the whole draft.

That second list is developmental editing, where chatbots are least reliable.

How much of a novel fits in ChatGPT, Claude and Gemini?

The context window is how much text the model can consider at once, measured in tokens. OpenAI's tokens help page gives the rule of thumb that one token is about three quarters of an English word, so a 90,000-word novel runs to roughly 120,000 tokens. Your prompt and the model's answer share the same window.

Plans, prices and limits as listed on chatgpt.com/pricing, the ChatGPT Plus help page, claude.com/pricing, the Claude context window help page and gemini.google/subscriptions on 19 October 2026. Training settings from each vendor's privacy help page on the same date.

Run the arithmetic for a 90,000-word draft. At about 120,000 tokens, it does not fit in ChatGPT Plus's 54K fast mode at all. It fits inside the 256K reasoning window on paper. But OpenAI's pricing page says the space for your input is smaller than the total window, puts the Plus reasoning input maximum at about 320 pages of text, and calls that figure approximate. A 90,000-word manuscript is in that range, so expect it to be tight rather than roomy.

Claude's paid plans and Google AI Pro list windows of up to 1M tokens, which would hold several novels. On the file side, OpenAI's file uploads FAQ caps text documents at 2M tokens per file and lists 3 uploads a day for Free users.

Why a big context window does not mean a good whole-book edit

A window that holds the book is necessary, not sufficient. Here is where chatbots break down on a full manuscript.

Losing track across chapters. Models pay uneven attention across very long inputs. Details buried in the middle third are the most likely to be missed. Ask "list every scene where the locket appears" and you may get four of the seven.

Flattery and agreeableness. Chatbots are tuned to be helpful and pleasant. Ask "is this good?" and you will usually hear that it is, with gentle suggestions. That is a poor substitute for a reader willing to say chapter 9 can go.

Rewriting your voice. Ask for feedback and many models hand back a "polished" version. It reads smoother and sounds like nobody in particular. If you paste that in, your style drifts toward a generic default, chapter by chapter.

No persistent scene map. A chat does not keep a structured record of your scenes, characters and timeline between sessions. Each new conversation starts from what you paste. Projects and memory features help a little, but they are not a scene-by-scene index.

Inconsistent repeat runs. Run the same "top five problems" prompt twice and you can get two different lists. That makes it hard to tell whether a revision fixed anything or the model just changed its mind.

None of this means you should stop. Treat the output as notes from a fast reader with a patchy memory.

Prompts that ask for diagnosis, not rewrites

These work in ChatGPT, Claude or Gemini. Each forbids rewriting and asks for evidence. Paste the chapter below the prompt.

1. Line notes without rewriting

You are a line editor. Do not rewrite any sentences. For the chapter below, list up to 15 specific issues (repetition, unclear antecedents, overused words, tangled syntax, dialogue tag clutter). For each, quote no more than 8 words of my text and explain the problem in one sentence. Leave the fix to me.

2. Honest scene diagnosis

Act as a blunt developmental editor, not a cheerleader. For this scene, answer: What does the point-of-view character want? What stands in the way? What changes by the end? If any answer is "nothing" or unclear, say so plainly. Do not praise the writing and do not suggest new prose.

3. Where a reader gets lost

Read this chapter as a first-time reader. Mark every point where you are unsure who is speaking, where we are, when this is happening, or why a character acts. Give paragraph numbers and a one-line reason. No rewrites.

4. Pacing check

Break this chapter into beats (one line each). Label each beat as action, dialogue, reflection, or description. Then tell me which stretch is slowest and why, in no more than 100 words. Do not suggest replacement text.

5. Consistency against your notes

Below is my story bible, then chapter 14. List every place where chapter 14 contradicts the story bible (names, ages, eye colour, timeline, who knows what). Quote the line and the bible entry. If you find nothing, say "no contradictions found" rather than inventing one.

6. Chapter ending strength

Look only at the final 300 words of this chapter. Does it end on an open question, a decision, a reversal, or a resolution? Would a reader turn the page? Answer in three sentences and give two ideas for raising the tension, described in plain terms, not written out as prose.

7. The devil's advocate pass

Assume this chapter has three serious problems. Find them and rank them by how much they would bother an agent or acquiring editor. Be specific and cite the text. Do not soften the assessment.

Run important prompts twice and only trust problems that appear both times. If the model slips into rewriting, reply "Diagnosis only. Remove all suggested prose."

How to run a whole-book pass with a chatbot

Do not paste 90,000 words and ask what is wrong. Break the job down so the model works from compact notes you control.

Step 1: Summarise each chapter. Paste one chapter per message and ask for a summary of no more than 150 words covering events, point-of-view character, goal, outcome, and any objects, promises or questions introduced. Save every summary in a document of your own.

Step 2: Build a running story bible. After each chapter, ask the model to list new or changed facts: names, ages, physical details, locations, dates, relationships, and what each character knows. Merge these into one bible file. Check it yourself, because errors here compound.

Step 3: Ask book-level questions of the summaries. Paste all the chapter summaries together. For a 30-chapter book at 150 words each, that is about 4,500 words, which fits easily in any plan. Now ask structural questions: where is the midpoint, which subplot vanishes after chapter 11, which setup never pays off.

Step 4: Verify against the text. Summaries lose detail, so treat every flag as a lead and confirm it in the chapter.

Step 5: Check continuity chapter by chapter. Use prompt 5 against the bible. Slow, but it catches more than one giant upload.

For a broader look at AI editing workflows, read how to use AI to edit your novel.

Is it safe to paste your unpublished novel into ChatGPT?

That depends on your settings, so check them before you paste anything.

ChatGPT. The pricing page says content on Free, Go, Plus and Pro is used to train models, with an opt-out. OpenAI's data controls FAQ says to go to Settings, then Data controls, and turn off "Improve the model for everyone." Temporary chats are not used for training but may be kept for up to 30 days for safety.

Claude. Anthropic's privacy help page says chats are used for training if you allow it through the Model Improvement setting, or if a chat is flagged for safety review. Incognito chats are not used, even with that setting on.

Gemini. Google's Gemini Apps privacy hub says activity can be used to train models while Keep Activity is on. Turning it off stops future chats being used for training unless you send feedback, but chats are still kept for 72 hours, and human-reviewed chats can be kept for up to three years.

If the manuscript is under contract, switch these settings off first.

Do you have to disclose ChatGPT editing to Amazon KDP?

Amazon separates AI-generated content from AI-assisted content. Getting notes on prose you wrote is different from pasting in text a chatbot wrote.

Our Amazon KDP AI disclosure policy guide covers what KDP asks and when. If you are unsure where your use falls, our explainer on AI-assisted vs AI-generated writing walks through the line. The diagnosis-only prompts above keep you firmly on the assisted side, because every sentence in the book stays yours.

General chatbot vs dedicated manuscript tool vs human editor

Chatbot prices as listed on each vendor's own page on 19 October 2026. The EFA 2026 rate chart lists a fiction developmental edit at 3.0 to 3.5 cents a word, which is $2,700 to $3,150 for 90,000 words.

A chatbot is the cheapest second pair of eyes on a scene. A dedicated tool saves the summary-and-bible legwork. A human editor tells you what the book is trying to be.

Where StoryWith AI fits

StoryWith AI is a manuscript analysis tool. You import the whole draft as .docx, .md, .txt or .pdf, and it reads what you wrote rather than writing for you. It runs analyzers for structure, setups and payoffs, pacing, dialogue, continuity and plot holes, theme, scene purpose and character arcs, then puts every finding into one ranked list linked to the exact scene. The features page lists what each one checks.

That covers the whole-book pass described above without the manual summaries. A book-level pass of a 90,000-word novel uses about 18 AI text actions, and importing it costs 5 more, once. The Free plan includes 50 one-time actions, and Author is $19 a month with 500 a month, per our pricing page. You can also test a chapter on simulated reader personas, which is a first pass, not proof of how real readers will react.

It is the wrong tool for some jobs. It does no line-level proofreading inside Word or Google Docs; ProWritingAid or Grammarly fits that better. For brainstorming, a general chatbot is more flexible. And it does not replace a human editor; it clears software-detectable issues first so the paid edit goes on judgment.

Frequently Asked Questions (FAQ)

Can I use ChatGPT to edit my novel for free?

Yes, at scene level. ChatGPT Free lists a 27K token window for its fast mode, with an input maximum of about 12 pages of text, so plan on roughly one chapter at a time. Free users also get 3 file uploads a day. A full novel needs the bigger reasoning window on a paid plan, or the chapter-by-chapter workflow above.

Is ChatGPT a good beta reader?

It is a useful simulated first reader for clarity and pacing on a chapter. It is not a substitute for real readers, because it tends to agree with you and has no genuine reaction to your story.

Is Claude better than ChatGPT for novel editing?

Claude lists context windows of up to 1M tokens depending on the model, so a full manuscript fits more comfortably. Both share the same long-book weaknesses, so prompt discipline matters more than brand.

How do I stop ChatGPT from rewriting my prose?

Open every prompt with "Do not rewrite. Diagnosis only." Ask for numbered issues with short quotes and one-line reasons.

Will ChatGPT steal or train on my novel?

On consumer ChatGPT plans, the pricing page says content is used for training unless you opt out. Turn off "Improve the model for everyone" in Settings, then Data controls, before pasting. Claude trains only if you allow it, and Gemini trains while Keep Activity is on.

Can ChatGPT find plot holes in my whole book?

Sometimes, but not reliably from one giant paste. Build chapter summaries and a story bible, ask about the summaries, and verify each flag in the text.