The restart problem changed as ChatGPT Projects gained built-in memory, files, instructions, app sources, and saved responses. A fresh chat inside the same Project is no longer necessarily a blank room. But “some relevant context is available” is not the same as “the exact approved book state is known.” The safe restart is a compact checkpoint plus a verification step.
A new chat is not always blank—but it still needs an authoritative checkpoint
Outside the original conversation or a shared Project, a new chat may have none of the book-specific material you expect. Inside the same ChatGPT Project, built-in project memory can reference other project chats and files, project instructions remain active, and saved responses can be reused as sources. That is real continuity and should be used rather than described away.
The remaining problem is authority. A Project may contain an early outline, three revisions of chapter six, a rejected character change, and a later correction. Relevant retrieval does not by itself identify which one is approved. The restart checkpoint supplies that decision: this is the accepted chapter, this is the current position, these facts are locked, and this is the next beat.
What to bring into the new chat
A clean restart hands the new conversation five things and no more: where you are, who matters, what must not change, the next beat to write, and how the prose should sound. That is enough to continue without contradicting yourself, and small enough that the assistant spends its attention writing rather than re-reading. The Restart Kit builder above assembles exactly these into a block you can paste.
- Position — the precise spot you stopped ("end of chapter 12; the siege just broke").
- Characters — the handful who are active, with the one detail each that a reader would notice if it changed.
- Locked facts — the canon a new chat is most likely to contradict: names, dates, a timeline, an injury.
- Next beat — the single scene or move you want written now.
- Voice — tense, point of view, and any hard rule about how it should read.
Restarting in a normal chat, a ChatGPT Project, or a durable book record
The Restart Kit works in every workflow. In a normal chat it re-supplies the missing context. In a ChatGPT Project it identifies which available records are authoritative. In a connected book workflow it becomes a verification summary generated from stored book state rather than reconstructed from memory. The difference is where the checkpoint comes from and how confidently it can be reproduced.
| Starting a new chat | Normal chat | ChatGPT Project | Connected book |
|---|
| Where you left off | Paste or restate it | Can be present in chats or saved sources | Read from the stored writing position |
| Accepted chapters | Paste the relevant text | Files, canvases, and responses may coexist | Retrieved as the accepted chapter version |
| Locked facts / canon | Only what you provide | Available if stored and retrieved | Held as approved blueprint and canon rules |
| Main risk | Missing context | Retrieving a superseded record | Stale status if an interrupted action was not reconciled |
| Restart action | Paste the kit | Open in the same Project and verify the kit | Select the book, reconcile status, and verify |
Projects materially improve continuity. A durable book record adds an explicit accepted version and current position when several useful records disagree.
A Restart Kit supplies missing context in a normal chat and establishes authority inside a Project. A durable book record can generate the checkpoint from accepted state.
The 60-second Project restart check before you draft
Open the new chat inside the same Project and ask for a status summary with five fields: current accepted chapter, next chapter to write, last locked plot or argument decision, active character or source constraints, and the response or file used for each answer. Do not ask for prose yet. Compare the summary against your own checkpoint and correct anything that points to an obsolete source.
Then name the exact source of truth for the session. If chapter nine in a saved response is approved and chapter nine in an uploaded draft is not, say that plainly or remove the obsolete file. Save the corrected status summary back to the Project so the next restart has a cleaner reference. Projects can preserve a living workspace, but the author still has to govern conflicting records.
Finally, request only the next bounded unit—a beat, scene, section, or chapter—and require the assistant to stop if the requested move conflicts with the checkpoint. That turns project memory into a useful retrieval layer without pretending it is an infallible manuscript database.
Give every piece of book context one job
A Project becomes reliable when you decide what belongs in each layer. Put permanent operating rules in project instructions: point of view, tense, audience, hard content boundaries, and the rule that uncertain canon must be questioned rather than invented. Put substantial reference material in files: the current outline, character canon, research notes, and the latest accepted manuscript export. Use project chats for working sessions and save only genuinely reusable decisions or summaries as project sources.
Do not store the same fact in five places without identifying which copy wins. If the character sheet says Mara is thirty-two, an early chat says thirty-four, and chapter three implies thirty, project memory has multiple plausible answers. The fix is not a longer restart prompt. Correct the authoritative character sheet, label or remove obsolete material, and state that the character sheet governs age conflicts. Retrieval works better when the workspace itself is coherent.
Keep changing session state separate from durable canon. “Write scene seven next” belongs in the current checkpoint because it will soon be obsolete. “Mara cannot read the inscription” belongs in canon because it must survive the whole book. Voice examples can live in a short reference file; the current approved chapter can live in a clearly named manuscript file or saved source. This division keeps the restart kit small while leaving the evidence available when the assistant needs it.
| Project layer | What belongs there | What does not |
|---|
| Project instructions | Stable writing and decision rules | A running chapter summary that changes daily |
| Files | Current manuscript, outline, canon, research | Several unlabeled revisions of the same chapter |
| Saved project sources | Approved summaries and reusable decisions | Every exploratory answer the assistant produced |
| Working chats | Brainstorming, revision, and bounded drafting | The only surviving copy of an accepted chapter |
| Restart checkpoint | Current position, next beat, active constraints | The complete history of the book |
Use a two-pass restart: reconcile first, draft second
The first message in the new chat should request a state check, not a chapter. Give the compact restart kit, name the authoritative files or saved sources, and ask the assistant to return a short reconciliation: current writing position, accepted version, active characters, locked facts, next beat, and any conflicts it found. Requiring the conflict list is important. A polished summary can hide uncertainty; an explicit discrepancy makes it possible for you to choose before new prose compounds the mistake.
Review that reconciliation against the manuscript. Correct wrong facts in plain language and identify which source should control the decision. If a stale upload caused the conflict, remove it or replace it instead of relying on a correction buried in one more chat. Once the returned state matches the book, give the drafting instruction as a second message. Limit the request to the next unit and repeat the non-negotiable constraints that are relevant to that unit.
At the end of the session, update the checkpoint from what was actually accepted, not everything discussed. Record the last accepted scene, any newly locked canon, the next intended beat, and any unresolved choice. A restart kit is valuable because it is concise and current. If it grows into a transcript, it stops being a checkpoint and recreates the same retrieval problem it was meant to solve.
- Open the fresh chat inside the correct Project.
- Paste or generate the current checkpoint and name the governing sources.
- Ask for a reconciliation and conflict list without requesting prose.
- Resolve every material conflict against the accepted manuscript or canon.
- Request one bounded writing unit, then approve or revise it.
- Update the checkpoint only with accepted changes before ending the session.
The verification message is not ceremony. It is the point where a cheap contradiction is caught before it becomes an expensive rewrite across several new scenes.
Worked example: three versions of the same chapter
Imagine a Project contains “chapter-08-draft.docx,” a later Canvas revision, and a saved response titled “Chapter 8 final notes.” The draft sends Lena through the north gate. The Canvas revision sends her through the river tunnel. The final notes approve the tunnel but preserve a sentence from the earlier draft. A new chat asked simply to continue may retrieve any relevant combination and write chapter nine as if both routes happened.
The restart checkpoint should say: “Accepted chapter 8 is the current Canvas version; Lena entered through the river tunnel; the north-gate route is rejected; chapter 9 begins when she reaches the archive stairs.” Then ask the assistant to identify whether any Project source conflicts with those statements. Once the conflict appears, rename or remove the obsolete draft and update the approved summary. The checkpoint solves the immediate restart, while cleaning the Project reduces the chance of the same mistake later.
If the Canvas is the working master, use its version history to inspect or restore an earlier revision rather than uploading multiple anonymous exports. OpenAI documents that Canvas can show changes, navigate versions, restore a version, and export writing documents as PDF, Markdown, or Word files. Export an accepted milestone when you need an external backup, but give the exported file a version or date so it cannot masquerade as the current manuscript six sessions later.
Recover when the new chat confidently remembers the wrong thing
Stop drafting as soon as a contradiction appears. Ask which chats, files, instructions, or memories supported the disputed statement, then compare those sources with the accepted manuscript. Current ChatGPT memory controls can expose sources used for personalization, but OpenAI notes that the displayed sources may not show every factor. Treat source visibility as diagnostic help, not proof that no other context influenced the answer.
Correct the durable source before continuing. Replace a stale canon file, remove an obsolete upload, move a misleading conversation out of the Project, or save a corrected decision as the Project source. OpenAI’s Project guidance specifically recommends removing older or unnecessary uploads, combining file data, or splitting work into multiple Projects when the material becomes unwieldy. That is a stronger repair than repeatedly telling one chat to ignore the wrong version.
If you cannot determine which record is right, fall back to the manuscript itself. Paste the relevant accepted passage, label it as authoritative for the session, and postpone any scene that depends on unresolved canon. The assistant should be instructed to flag unknowns instead of filling them. A momentary pause preserves continuity better than confident invention that forces revisions across later chapters.
| Failure | Immediate containment | Durable repair |
|---|
| Wrong character fact | Restate the accepted fact and pause the scene | Correct the governing canon file and remove conflicting copies |
| Continues from an old chapter | Name and supply the accepted chapter | Replace or relabel stale manuscript versions |
| Voice drifts | Reapply the active voice rules and sample | Put stable rules in project instructions and a concise reference |
| Invents an unresolved decision | Reject the invention and mark the choice open | Add an explicit unresolved-decisions section to the checkpoint |
| Project is too noisy | Use only named governing sources for the session | Remove, combine, or separate obsolete material |
Prepare the handoff before a long chat becomes unusable
The best restart begins before the old conversation ends. When a session reaches a meaningful milestone, ask for a handoff summary, then verify it yourself. It should name the last accepted text, not merely the last generated text; separate locked canon from suggestions; record the exact next writing position; and list unresolved decisions. Save the corrected response as a Project source when it will be useful again, a capability OpenAI provides directly inside Projects.
Preserve the prose outside the chat. Canvas can export documents as PDF, Markdown, or Word and provides version navigation and restoration, while a general ChatGPT data export includes chat history and other account data. Those are useful archives, but a full account export can take time and is not a practical manuscript versioning workflow. Keep regular, named manuscript exports in storage you control, particularly at chapter, part, and full-draft milestones.
Be deliberate about deletion. OpenAI says chats remain in an account until manually deleted; archived chats remain retained, while deleting a chat removes it from visible history and schedules deletion under the applicable retention policy. Project files are retained until the Project is deleted under the documented policy. Archive when you only want a chat out of the sidebar. Delete or move obsolete material when its continued presence is causing retrieval conflicts, after preserving anything you actually need.
- Name the last accepted chapter, scene, or paragraph and where its master copy lives.
- Separate approved canon from ideas that remain optional.
- Record the next beat and the first decision the next session must make.
- List active voice, point-of-view, timeline, and research constraints.
- Remove or clearly label the obsolete records most likely to be retrieved by mistake.
- Export an accepted manuscript milestone outside the conversation.
The version where the project remembers, not you
When a book lives in BookWriter, a new conversation does not start in a blank room. It selects the saved project, reconciles the current status, loads the active chapter packet, and continues from the stored cursor. You did not paste anything; the project told the assistant where things stand. That is the whole point of connecting an assistant to a durable book system — the conversation stays disposable, and the book does not.