How to keep your story consistent across Claude and ChatGPT chats

You ask Claude to check a chapter for inconsistencies. It finds one: a character who says he has never been inside the house knows where the spare key is.

Unfortunately, that was the clue.

Before asking an assistant to check continuity, give it a way to distinguish a mistake from something you put there on purpose. That becomes especially useful when you start a new chat or move a scene from ChatGPT to Claude.

A small story bible can do that. It is a reference for what you have established, what remains undecided and what each character knows. It does not need to contain your entire manuscript.

Here is a setup to try with one chapter. The story, excerpts and expected findings below are invented examples, not a transcript of a live product test.

Start with five notes, not another writing project

For the example, we will use a mystery called The Spare Key. Nora is clearing out her late aunt's house. Eli, a neighbour, offers to help. He says he has never been inside. Later, he tells Nora where to find the spare key to the locked garden shed: in a tin inside the house.

Keep the reference small enough that you are willing to maintain it.

The filled example shows all five. There is also a blank Markdown version for your own story. Use them as sections in one file or as separate notes.

Give each important detail one main home. Birth dates belong in Characters. The scene date and planned reveal belong in Timeline. Other notes refer there rather than maintaining their own copies.

Separate the truth from what people believe

Eli's statement is not a fact about the house. It is something a character says.

The author has established that Eli visited the house in 2007. Nora has not learned that yet. The reader has heard his denial and is about to see a reason to question it.

Keep an idea such as “perhaps Eli stole something” separate. It is a possibility, not an explanation the assistant should quietly add to the scene.

This distinction matters during revision too. A note saying “Eli has never entered the house” loses the fact that he is lying. Write “Eli claims he has never entered the house” and keep the actual history alongside it.

For uncertain details, “not established” is a useful answer. You do not need to invent a motive just to fill a field.

Tell the assistant what not to fix

In Intentional clues, explain the apparent contradiction:

Apparent contradiction: Eli denies entering the house in chapter 2, then knows where the spare shed key is in chapter 5.
Purpose: Give Nora and the reader a reason to question him.
Preserve: Both the denial and his knowledge of the key.
Do not reveal yet: The earlier visit, before the planned reveal in Timeline. His motive is still undecided.

That is more useful than a general instruction to “keep the mystery subtle”. It names the thing you want protected and the information you are not ready to reveal.

Put the reference where your assistant can read it

With files: download the filled example and attach it to your chat. In the review prompt below, replace the opening Hjarni sentence with “Read all five sections of the attached Story Bible.” At the save step, edit the file yourself and attach that updated version to the next chat.

With Hjarni: connect your assistant using the ChatGPT or Claude guide. Check access before starting: Hjarni's ChatGPT guide currently lists Plus, Pro, Business, Enterprise or Education; its Claude guide lists Free with one connector as well as paid plans. Workspace restrictions may also apply. The example needs five available notes; Hjarni's free plan allows 25.

Then send this setup request:

Hjarni, read https://hjarni.com/templates/story-bible and create its “Story Bible (example)” folder, folder instructions and five example notes. Check first: if that folder already exists, stop and tell me. Do not add notes from this conversation, customise the example with my own material or change any other folder. Link to the folder and list the notes you created.

Review any approval request. Check that the folder has the five notes and its instructions before continuing. If setup stops partway through, inspect what was created before retrying rather than starting another copy.

Review a scene before asking for a rewrite

With the reference in place, try:

Hjarni, read all five notes in “Story Bible (example)”: Story overview, Characters, Timeline, Intentional clues, and Ideas and decisions. Tell me which notes you read. If you cannot retrieve one, say so before reviewing. Check the scene below against them. Separate likely mistakes, intentional clues and details the notes do not resolve. Point to the relevant note and section for each finding. Do not rewrite the scene or edit any notes.

Follow it with this deliberately flawed scene:

“The spare shed key? Top shelf of the pantry,” Eli said. “In the blue tin.”

At thirty-six, Nora still needed the stool to reach it. She pulled the tin down and found the key inside. She remembered Mara telling her about Eli's visit in 2007.

“How did you know it was there?” she asked.

The example notes give Nora's birth date as 17 June 1988 and the scene date as 12 October 2025. A useful review should flag her age as 37, not 36. It should also flag her memory of the visit: she has not learned about it at this point.

It should leave Eli's knowledge of the key alone. That is the clue. It should not conclude that he stole anything, because the notes have not established that.

It may raise other points too. Treat those as questions to check, not automatic errors. Look for findings that point to something specific in the reference.

Save the change you actually approve

Suppose you decide to reveal Eli's visit later. Move the planned reveal from chapter 9 to chapter 11, while keeping the clue in chapter 5.

With Hjarni, ask:

Hjarni, read Timeline in “Story Bible (example)”. I have approved moving the planned reveal of Eli's 2007 visit from chapter 9 to chapter 11. Keep the chapter 5 clue and all established events unchanged. Update the current reveal entry and record chapter 9 as superseded in the revision section. Do not edit the manuscript or other notes. Read the saved note back and link to it so I can check the change.

With files, make that change in the Timeline section yourself. Keep chapter 11 in the current reader-order table, and mention chapter 9 only in the revision entry as the old plan. Save the file; a reply in the chat does not update your local copy.

Check the saved reference. The reveal should have moved, not the visit itself or the clue. In Hjarni, you can inspect note history to review the edit. Neither path needs a manuscript rewrite.

Come back in a new chat

Open a new conversation with Hjarni available, or attach your updated Markdown file. To try the handoff between assistants, use the other one with access to the same Hjarni folder or the same updated file.

Do not put the revised chapter number in your question:

Hjarni, read the current notes in “Story Bible (example)” before answering. When is the reader supposed to learn about Eli's earlier visit? What has Nora learned by the end of the key scene? Have I established why Eli lied? Link to the notes you used. Do not change anything.

For the file route, replace the opening sentence with “Read the updated Story Bible attached to this chat before answering.”

Look for the current reveal plan, Nora's limited knowledge after the key scene and the still undecided motive. If the answer gives the old plan or invents a motive, inspect the reference the assistant actually used.

A correct answer alone does not show where it came from. For a stronger check, change the reveal chapter directly in Hjarni without telling the assistant the new value. Start another chat and ask it to retrieve Timeline before answering. Check the retrieved note as well as the answer. This checks whether the saved update is being used in that run, not whether every future answer will be correct.

Where Hjarni fits

You can try the method with a Markdown file. Hjarni is useful when you want Claude and ChatGPT to read and update the same saved story notes, rather than keeping separate copies for each chat.

Keep the manuscript wherever you already write. This workflow updates the reference in Hjarni, not an uploaded Project file or the manuscript in another application. Connecting Hjarni does not make every brainstorm part of the reference, and it does not guarantee that every answer will consult it. Ask for the relevant notes explicitly.

Use a chat where the connection and required actions are available. OpenAI's app guidance explains account access and approvals. Hjarni does not use your notes to train AI models. The assistant you connect handles what it reads under its own provider's policy.

For a real writer's experience, see how Jim uses shared story notes with ChatGPT and Claude. Keep voice preferences separate from plot facts; the Writing Style Guide is a companion for that job.

Try it with your own story

Download the blank Story Bible and start a separate reference for your story. Leave the fictional example where it is. Two characters, the events around one scene and one thing an editor must not “correct” are enough.

The Story Bible template page includes a setup prompt for creating those five notes in Hjarni using your own material. Ask for a review, check the findings and save only the changes you approve. You do not need to catalogue your whole novel to find out whether this helps.