← sources

How do you keep state/context between chats in a long Claude project without the chat closing eating your usage?

source: reddit (reddit:ClaudeAI)published 2026-11-01 12:29ingested 2026-11-02 02:38

retained text

I use Claude (Pro) as a job search partner in one Project. State lives in editable docs (artifacts), not in the chats:

* HQ doc**:** current status, next step, open items, standing decisions, and a map of the other docs. Every new chat reads only this.
* Job-specific docs: master profile, CV, cover letter kit, an answers doc for unsent applications, and an archive for sent ones etc. etc.
* Rules in the project instructions**:** one job per chat, Sonnet for admin, Opus only for applications, and Claude tells me when a chat is getting heavy.
* Claude Code runs a cheap daily job scan and logs what I sent or skipped (career ops ftw).

At close, Claude rewrites the HQ “now” and “next step” lines, files finished work in the right doc, updates my tracker and gives me a prompt for Claude Code. So far this has worked great in the sense that I was able to keep context across different chats in the project without much manual work.

The problem: last closing request cost me over 5% of my usage. The chat had run all day, and the close alone took about 18 tool calls, each re-reading the whole chat. In general I noticed that even those steps eat much more of my usage. 



Since this is surely something many of you have ran into, i'd like to get your insights into:

1. Has anyone found a cheaper way to close or hand off?
2. Do doc connector results count a lot toward usage, compared with cached project knowledge?
3. How long do you let a chat run before cutting/closing it?

Thank you!

document fetched 2026-11-02 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xl1uyx/how_do_you_keep_statecontext_between_chats_in_a