← sources

Compaction wiped out the continuity in my 670-turn co-writing chat. Here's how I got it back with subagents and a checkpoint doc.

source: reddit (reddit:ClaudeAI)published 2026-10-26 12:58ingested 2026-10-27 02:36

retained text

I co-write a long romance novel with Claude, one turn at a time. I write one character's lines and Claude handles everything else. One chat got to roughly 670 turns, compaction kicked in, and Claude started forgetting things it had written itself.  
The plot survived. The texture didn't. A character's career history got scrambled. Things one character had already confessed got "revealed" a second time. A piece of clothing that went home with one person on Monday turned up on the other person's chair on Tuesday. Each slip was small, but together they made a book that didn't trust itself.  
**The fix:** I had Claude split the old chat into six slices of about 112 turns each. It sent one subagent into each slice at the same time, using the past-chat reading tool. Every agent filled in the same six-part template:

Scene-by-scene log

My character's dialogue, word for word, labeled by turn  
What each character has already told the other  
Canon facts (names, dates, places, who has what)  
My corrections and rules, quoted exactly  
Open threads and unresolved tensions  
Then the main chat merged the six files into one checkpoint doc and checked it against my character bibles, with the bible winning any conflict. I read the result and corrected it before treating it as canon. Agents summarize, and summaries can be confidently wrong.  
Sections 3 and 5 did the most work. "What has she already said out loud" is exactly what compaction loses. Your corrections are the rules you'd otherwise have to repeat forever.  
**How I keep it from happening again:**  
Say "update checkpoint" after every scene. Small and frequent beats big and rare.  
Keep the character bibles as separate Project docs and call them canon.  
Put standing rules in Project memory or instructions, not just in the chat.  
Start a fresh chat before the old one gets huge, and paste a recovery prompt first.  
Keep separate stories in separate Projects so their timelines don't bleed together.  
**Caveats:** this needs "search and reference past chats" turned on, and it can only see chats in the same Project. It's also slow and uses a lot of tokens on a long chat. So the boring answer is the best one: checkpoint often. Two minutes of checkpointing saves twenty minutes of rebuilding.  
I put the full workflow on one page, including the agent brief and the recovery prompt with copy buttons: [https://claude.ai/artifact/Sxmz51DY2wC8XsrVwnSRoD](https://claude.ai/artifact/Sxmz51DY2wC8XsrVwnSRoD)  
(Bonus: while Claude was writing the original version of this post, the chat compacted. The tutorial got eaten by the thing it was about.)

document fetched 2026-10-27 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xfw3kv/compaction_wiped_out_the_continuity_in_my_670turn