Every message re-sends the whole conversation, so a long one uses up limits fast and recalls its early details worse. A new conversation that starts from a short handoff costs a fraction and stays sharp.
When to offer it
- The user asks to start over, hits a limit, or says replies got slow or forgetful.
- The conversation is clearly long: many long turns, large pasted files or documents. Offer it once, in one sentence, at a natural break. Don't interrupt a short conversation or a task that is nearly done.
The handoff
Write it in the user's language, under 300 words, inside one code block so it copies cleanly:
- Goal — one line.
- Decisions so far — bullets, each with its reason in a few words.
- Current state — what exists now: files, drafts, versions, results. Quote only short snippets that the next step needs; name everything else.
- Next step — the very next thing to do, then any open questions.
- Must not be lost — names, numbers, constraints, preferences, things that were tried and failed.
Then give the saving in one line, as an estimate: "This conversation is about 60K tokens; the handoff is about 500, so each message in the new chat re-sends under 1% of what it does here." Then tell the user: open a new conversation, paste the block, and continue.
On each surface
- claude.ai chat: a new chat with the pasted handoff. In a Project, the project's files stay available, so the handoff can just name them.
- Cowork: start a new task with the handoff as its first message.
- Claude Code:
/compactkeeps working in place and is usually enough. For a clean break,/clearand paste the handoff./cco task patchkeeps structured task state across/compactautomatically.
Never paste whole files, logs or the transcript into the handoff.