Your Claude session is lying to the next one
The work may be done, but the handoff is broken. Here’s the closeout ritual that keeps decisions, file changes, memory updates, and review warnings from disappearing into the transcript.
A Claude Cowork session can look productive and still leave you with a bad handoff.
Claude drafts the memo.
It edits the spreadsheet.
It helps organize a folder.
It pulls a scattered pile of notes into a useful brief.
Then the session ends, and the useful state gets trapped inside the transcript.
Tomorrow, you’re back in the same project asking questions you already paid the model to answer.
Which file is current?
What did Claude change?
What still needs review?
Which assumption was weak?
What should the next session know before it starts?
That’s the leak here.
The work happened, but the handoff didn’t.
For normal chat, this is annoying.
For Cowork, it’s operational drag because the system is built around work that can touch files, apps, browser context, connected tools, scheduled runs, and multi-step deliverables.
Anthropic’s Cowork safety guide says users control which local files Claude can access, and that granted access can let Claude read, write, and permanently delete those files. The same guide recommends dedicated working folders and backups instead of broad access to sensitive directories.
That changes the meaning of “done.”
A real Cowork run shouldn’t end when Claude stops responding.
It should end when the next session has enough state to continue without making you rebuild the whole job from memory.
The best clue came from the skills crowd
The strongest recent pattern I found wasn’t another fancy skill.
It was a closeout skill.