Creative writing
How to Use ChatGPT to Write a Novel
A novel is not a long prompt. It is a sequence of choices about desire, conflict, voice, causality and meaning. ChatGPT is useful when it makes those choices easier to inspect: it can challenge an outline, generate alternatives, track a cast list and suggest revisions. The writer remains the source of intent and the final judge of language.
GPTPrompts.AI Editorial
Hands-on with ChatGPT; every method here is one we use · Last updated October 2026
Affiliate link, we may earn a commission at no extra cost to you.
01 · Reality check
What ChatGPT actually does well here
Good at
- Turning a premise into alternative outlines
- Maintaining visible character and continuity tables
- Generating scene questions and revision checklists
- Offering line-edit options on supplied prose
- Finding contradictions and missing transitions
Not the right tool for
- Replacing the writer's point of view or lived knowledge
- Guaranteeing originality, quality or publication
- Doing research without source verification
- Remembering canon that was never recorded
- Deciding what the novel should mean
02A · Working notes
Write a creative brief before asking for chapters
Record the premise, genre, intended reader, emotional promise, point of view, tense, setting, content boundaries and what the book must not become. Add three works you want to learn from only as high-level references, not as a request to imitate a living writer. State the decision you want from ChatGPT: premise options, conflict escalation, scene beats or a critique. A brief keeps the session attached to your project rather than producing generic prose. Keep the document outside the chat as your source of truth and update it when your story changes.
Build a story bible that distinguishes fact from possibility
Create tables for characters, relationships, goals, fears, secrets, changes, setting rules, timeline, objects and unresolved questions. Label each row CANON, OPTION or OPEN. Ask ChatGPT to repeat the canon before suggesting anything and to flag contradictions instead of silently fixing them. Include the source scene or note for a fact when the project is large. Do not ask the model to remember a rule that exists only in a previous conversation. A visible bible makes continuity review possible and lets you reject a plausible suggestion that does not belong in your book.
Outline by change, not by events
For each scene, state the viewpoint character, immediate want, opposition, new information, choice, emotional turn and exit condition. Ask ChatGPT for three possible scene shapes, then choose one and explain why it serves the arc. Track what changes in the reader's knowledge and the character's situation. A list of events is not an outline if the characters end each scene in the same state. Keep a scene ledger with chapter, location, time, characters, continuity dependencies and a short purpose. This turns a long manuscript into reviewable units.
Draft in small, controlled units
Give ChatGPT the scene goal, canon facts, sensory anchors, forbidden inventions, approximate length and the kind of help you want. Ask for beats or a rough passage rather than an entire novel. You can paste your own draft for questions, alternatives or a line-level edit, but say what must remain unchanged. If a detail is missing, require [PLACEHOLDER] rather than a fabricated street, recipe, medical fact or historical claim. Save your version separately and treat the output as material to accept, reject or rewrite—not as a final chapter.
Protect voice and avoid bland prose
Give the model a short sample of your own prose and describe its observable features: sentence length, distance, humour, imagery, dialogue density and taboo phrases. Ask for a diagnosis and options, not a request to copy a named author. Compare any edit against your original for rhythm, specificity and point of view. Generic intensifiers, symmetrical paragraphs and explanatory dialogue are signs to inspect. The most useful edit may be a question such as 'where does the scene state what the reader already understands?' Keep the sentence that sounds like you, even when the model offers smoother wording.
Use research as a source-led workflow
For history, science, law, geography, culture or lived experience, make a research list before drafting. Ask ChatGPT to identify questions and search terms, then verify facts in primary or authoritative sources. Record source, date, scope, confidence and where the fact enters the manuscript. Do not use generated citations as evidence. If you invent a fictional institution or technology, mark it as fictional and keep its rules consistent. Sensitive communities and experiences require human reading, sensitivity review and respect; a model's fluent description is not lived knowledge.
Run continuity and causality passes separately
After a section is drafted, ask for a table of names, ages, dates, locations, objects, injuries, knowledge states and promises. Then run a second pass for cause and effect: what does each character know, why do they act now, what prevents the obvious choice and what consequence carries forward? Do not ask for a general 'is this good?' verdict. Ask for evidence-backed flags with scene references and severity. You decide whether a flag is a deliberate mystery, a mistake or an opportunity. Keep accepted corrections in the story bible.
Revise at three scales
Use separate prompts for structure, scene and line. Structural review asks whether the central desire escalates, reversals earn their place and the ending answers the book's promise. Scene review asks whether each beat changes something and whether the viewpoint remains stable. Line review asks about clarity, rhythm, repetition and image precision. Mixing all three creates noisy feedback and encourages cosmetic edits before the story works. After each pass, make the changes yourself or approve them in a tracked copy. A clean file is not proof that the right changes were made.
Keep authorship, privacy and rights visible
Do not paste private correspondence, unpublished work belonging to another person, client material or identifiable beta-reader comments without permission. Review the current account controls and the terms that apply to your use. If you are submitting to a publisher, contest, school or client, check its disclosure and AI policy before using generated text. Keep drafts, prompts and major decisions in a dated project folder. ChatGPT can help you think, but you should be able to explain what you wrote, what you accepted and what you changed. Never promise that a generated passage is unique or automatically safe to publish.
A repeatable chapter workflow
Begin with the chapter's promise and the scene ledger. Ask for an outline that lists viewpoint, want, opposition, turn and consequence. Choose the beats, draft one scene and save it. Run a continuity table, then revise the scene yourself. Ask for a skeptical reader's questions and mark which are answered later. At the end of the chapter, update canon, timeline, open questions and research tasks. Once several chapters exist, ask for patterns across the ledger rather than asking the model to rewrite the manuscript in one pass. Small loops preserve your agency and make mistakes cheap to correct.
Use ChatGPT's file workflow deliberately
OpenAI documents support for uploading documents and asking ChatGPT to analyse their contents, with limits and availability that vary by plan, account and workspace. Upload only the material the task needs, and keep a local or approved copy of the manuscript. Ask the model to identify which file and section it used. For a long book, work by chapter or a curated packet and maintain a version label; do not assume that a chat has perfect access to every previous page. Review important continuity details against the manuscript itself. The file feature is a convenience, not a substitute for version control.
End with a human editorial decision
Before calling a chapter done, read it without the chat open. Check whether the voice, emotional movement, facts, pacing and promises belong to your novel. Ask a trusted human reader for reactions that a model cannot supply reliably, especially about cultural specificity, harm and emotional truth. Then record what you changed and why. The final question is not whether ChatGPT approves the prose; it is whether the work says what you mean and earns the reader's attention.
Design a character arc as a sequence of choices
For each major character, write the starting belief, desire, fear, strategy, relationship pressure and possible change. Ask ChatGPT to identify choices that would test the belief, but do not ask it to announce the arc in the manuscript. A strong arc is visible through decisions, consequences and revised behaviour rather than a paragraph explaining what the character learned. Keep the table labelled CANON or OPTION. If the model proposes a dramatic turn, test whether earlier scenes prepare it and whether the character has a reason to choose it. Reject an exciting idea when it breaks the emotional contract. The writer decides the meaning of the change; the assistant can make alternatives easier to compare.
Plan pacing with a chapter ledger
Track chapter purpose, viewpoint, scene count, approximate time, new information, reversal, unresolved question and emotional temperature. Ask ChatGPT to find long stretches where nothing changes, repeated exposition or a mystery answered before the reader has a reason to care. Require scene references and a proposed question, not a command to rewrite the book. Pacing is not the same as making every paragraph shorter. A quiet scene can earn its length when it changes a relationship or creates a choice. Read flagged passages yourself and record which change you accepted so a later edit does not restore the problem.
Write dialogue that carries action
Give ChatGPT the scene goal, each speaker's current knowledge, what each one wants to hide and the relationship history that matters. Ask for a subtext table or three possible exchanges, not dialogue in the voice of a named writer. Check every line for what it changes: information, power, commitment, misunderstanding or emotional temperature. Models often make characters too cooperative and make everyone explain the same context. Ask for a version with fewer explanations and mark any line that sounds like an author briefing the reader. Then rewrite in your own ear, read the scene aloud and check that the physical action is possible.
Use beta-reader questions instead of a synthetic verdict
A useful reader questionnaire asks where attention increased or dropped, what the reader believed each character wanted, which transition was confusing, what promise they expected the ending to answer and which detail stayed with them. ChatGPT can help turn your aims into questions or group feedback, but it cannot be your only reader and should not manufacture a consensus. Paste anonymised feedback only when you have permission. Keep each reader's voice separate before clustering themes. Ask the model to show the underlying comments for each theme and mark uncertainty. You decide which feedback belongs to the book and which would make it a different book.
Prepare a submission or handoff package
When the manuscript is ready for an editor, agent, contest or publisher, ask ChatGPT to check the package against the current guidelines that you supply. It can format a synopsis outline, query checklist, character list or version manifest, but do not let it invent credentials, publication history, endorsements or market claims. Verify word count, category, sample pages, file format and disclosure requirements yourself. Keep the manuscript, synopsis and query in a dated folder. Review the package for consistency with the actual book. A clean submission is accurate, follows current instructions and represents the writer honestly.
Recover when a draft goes in the wrong direction
When a generated scene feels generic or breaks canon, do not keep adding style instructions to the same output. Save the failed version, identify the failure in concrete terms and return to the last approved story-bible entry or scene ledger. Ask ChatGPT to diagnose the mismatch, list the canon it used, and propose three smaller repairs. You might need to restore a missing constraint, change the viewpoint, supply the character's immediate choice or remove a false research detail. Choose the repair yourself, update the source-of-truth notes and draft again in a smaller unit. A visible failure is useful feedback; hiding it inside a longer prompt makes the project harder to control.
02 · The method
Step by step
- 1
Write the brief
State premise, reader, genre, voice, boundaries and the decision you want help with.
- 2
Create the story bible
Label canon, options and open questions for characters, setting and timeline.
- 3
Plan scene change
Define viewpoint, want, opposition, turn and consequence before drafting.
- 4
Draft a small unit
Use supplied facts and placeholders; save the generated material separately.
- 5
Run targeted reviews
Review structure, scene movement, continuity and line separately.
- 6
Verify research
Open primary or authoritative sources and keep a fact ledger.
- 7
Protect rights and privacy
Check permissions, publisher rules, account controls and sensitive material.
- 8
Make the editorial call
Read the work yourself and accept only changes that serve your novel.
03 · Use this now
Copy-paste prompt for a novel scene
Act as a careful fiction development partner. Project brief: [premise, genre, intended reader, point of view, tense, voice and boundaries]. Canon for this scene: [facts]. Scene goal: [what must change]. Viewpoint character wants: [want]. Opposition: [force]. Required turn and consequence: [details]. First return a beat outline and list any missing information. Do not invent canon, research, names, dates or technical details; use [PLACEHOLDER] when needed. After I approve the beats, draft only [length] words. Preserve the viewpoint, avoid imitation of a named living author, and return a short continuity checklist plus three revision questions. Treat every suggestion as optional and do not decide the story's meaning for me.
04 · Avoid these
Common mistakes
- Asking for a whole novel before establishing a story bible
- Letting plausible invented details become canon
- Flattening your voice with generic line edits
- Using generated research or citations without checking sources
- Uploading private or third-party material without permission
- Treating a model's originality or quality claim as a publication guarantee
05 · Questions
Frequently asked questions
Can ChatGPT write a whole novel for me?
It can generate drafts and ideas, but a useful novel still needs the writer's premise, choices, voice, revision and judgment. Small, reviewable scene loops are safer than one giant generation.
How do I keep ChatGPT from forgetting my characters?
Maintain a story bible with canon, options, timeline and source scenes. Give the relevant excerpt to each session and ask ChatGPT to flag contradictions instead of filling gaps.
Can ChatGPT imitate a living author?
Do not ask for direct imitation. Describe observable craft features and use your own voice, public-domain references or broad genre conventions as the brief.
Should I upload my manuscript to ChatGPT?
Only after checking account controls, applicable terms, privacy needs and any publisher or client requirements. Upload the minimum necessary material and keep a versioned copy outside the chat.
Can ChatGPT verify historical details in fiction?
It can suggest research questions, but verify details in primary or authoritative sources and keep a fact ledger. A fluent answer or generated citation is not evidence.
Related guides
Primary sources
- OpenAI Help: Uploading files and audio to ChatGPT
- OpenAI Help: Using Library to manage files in ChatGPT
- OpenAI Terms of Use
Product menus and plan limits change. The linked vendor documentation is the authority when your screen differs.