What to do when a Claude Code conversation gets too long

Long Claude Code conversations get summarized again and again, and details fade. Here is what compacting does, what starting over loses, and how to carry your work into a fresh conversation with a note you can read.

Work on one thing in Claude Code for a few days and the conversation gets long. Nothing breaks. Claude just starts asking about things you settled yesterday, or brings back an approach you already dropped.

What Claude Code does on its own

When a conversation no longer fits, Claude Code summarizes the earlier part and keeps going. This is compacting. It happens automatically, and you can also run /compact yourself.

It works, and most of the time you do not notice it. But each summary is a summary of a summary. Which of your decisions made it through, and which did not, is hard to see from where you sit.

The other built-in choice is /clear: a new, empty conversation. Nothing fades because nothing comes along. You explain everything again.

The middle path: a handoff note

What people do by hand is ask Claude to write down where things stand, start a new conversation, and paste that in. It is the right idea. It also takes a few minutes, and the note is only as good as the request you typed that day.

Alien Desk's Warp does this in one click, from the tab's … menu:

• It writes a handoff note from the conversation: what was done, what you decided, what is still open.
• It opens a new conversation in the same tab, in the same folder, on a clean screen.
• The new Claude reads the note, tells you in a few lines what it picked up, and waits for you.

It takes about 30 seconds to a minute.

What goes into the note

The note always has the same ten parts, in this order:

• When the conversation was last active
• What this is: the project and the goal
• Names and numbers: folders, branches, commits, versions, IDs, copied exactly
• Current state, each item marked done or not done
• What you decided, with your reason
• What is waiting for your answer or action
• What not to do
• What was tried and did not work
• Next steps, in order
• What has not been verified

Two rules matter most. A decision counts as yours only if you said it in your own words; an idea you floated, or something Claude recommended, is written down as open. And names and numbers are copied as they appear, not rewritten.

The note is written in the language you were using.

Why the new Claude speaks first and waits

A note can be out of date the moment it is written. So the first request in the new conversation asks Claude to read the note, check it against the actual files, say what it picked up and what it plans to do next, point out anything that does not match, and then stop.

In our tests this was the part that mattered most. In one conversation the note said the next step was to upload a build. The new Claude read the project's own records, found the build had already been submitted, and said it would not do that step. If you answer only "go ahead", it does the first step that needs no approval and nothing more.

What stays and what changes

• The old conversation is not deleted. You can open it again with claude --resume.
• The tab's name, notes and pin in Alien Desk follow you to the new conversation.
• The terminal screen starts clean.
• The note is a file you can open, in ~/.claude/alien-desk/warp/. Only your account can read it, and notes older than 30 days are removed.
• A telepathy link follows to the new conversation.

When Alien Desk suggests it

Warp never runs by itself. After a conversation has been compacted three times, the alien on that tab asks once whether you want to warp. Say no and it does not ask again for that conversation. Across our own 88 conversations, one had reached that point.

You can warp at any time from the menu, for example right after finishing one piece of work and before starting the next.

How we tested it

• We took five real long conversations, had the note written from each, and gave only the note to a new Claude with questions about the work: next steps, decisions, things not to do, exact identifiers, and trick questions about things that were never said. On the three conversations we scored in every round, the shipped version got 34 of 36, with no answer that would have caused harm if trusted. It did not invent answers to the trick questions.
• Then we had a new Claude actually pick up each of the five, with read-only access to the real project. All five stopped and waited after their first reply. None tried to release, delete or pay for anything, and none went back to an approach that had been dropped. A separate Claude judged three ready to continue as they were and two as needing one or two things filled in.
• We also tried letting a Claude that had the whole conversation in context write the note. It took two to three times longer, read about fifteen times more, and scored no better, so the note is written from the end of the conversation plus the most recent summary.

Limits

• The note can leave things out. Reasons behind a decision and side topics from the middle of a conversation are the most likely to go missing. The old conversation is still there when that happens.
• It works in Terminal app tabs for now, not yet in iTerm2 or Ghostty.
• The tab has to be resting. You cannot warp while Claude is working.
• Writing the note uses a little of your own Claude usage.
• The conversations we tested on were our own, and our projects keep handoff documents. We do not know yet how well it does on a project with no written records.

Questions people ask

**Isn't this just /compact?** Same root: both summarize. Compacting keeps you in the same conversation with a summary you do not see. Warp starts a new conversation with a note you can open, at a moment you choose, and the new Claude checks the note against your files before it does anything.

Does Claude forget everything else? It starts from the note, your project's files, and your own instruction files such as CLAUDE.md. It does not see the old conversation.

Can I go back? Yes. The old conversation is untouched.

• Warp is part of Alien Desk Pro, free to try for 14 days.