Move a long chat to a fresh one without losing your place

You're two hours into a chat with ChatGPT or Claude. Early on you said "no bullet points" and "keep it under 200 words". Now the answers come back in bullets, run long, and suggest the supplier you ruled out an hour ago.

Most people do one of two things. They keep pushing in the same chat, correcting it again and again. Or they open a new chat and retype what they remember, and lose the decisions and exact wording that took an hour to get right.

This guide gives you one prompt that packs the work into a single block, the handoff, so a fresh chat can pick up where the old one stopped. It takes about two minutes.

Why long chats go wrong

Researchers at Microsoft Research and Salesforce tested more than 200,000 conversations across 15 leading models (Laban et al., "LLMs Get Lost In Multi-Turn Conversation"). When the details arrived over several turns instead of in one complete prompt, performance dropped by 39% on average.

In the early turns, the models guessed at what was missing and rushed an answer. Later turns built on those guesses, and once a model went off track, it rarely recovered. A fresh chat that starts from one summary you've checked doesn't carry those guesses with it.

Heads up: On paid Claude plans with code execution turned on, Claude summarises earlier messages by itself when a chat nears its limit. That keeps the chat going, but it's still the same chat, guesses included. The handoff gives you a fresh start that you've checked yourself.

The handoff prompt

Paste this at the end of the chat you want to move.

Write a handoff so I can continue this work in a new chat. Output one text block and nothing else.

Do not retell the conversation message by message. Capture the state of the work.

Include:

Goal: what I'm trying to achieve, in one sentence.
Status: what's done, what's in progress, what's stuck.
Decisions: what we decided and why. If something changed along the way, keep only the latest version.
My rules and preferences: copy my exact words. Don't paraphrase.
Exact details: names, numbers, dates, links, files, error messages. Refer to files and links by name instead of copying their contents.
Rejected ideas: one line each, with the reason, so the new chat doesn't suggest them again.
Open questions.
Next step.

Mark anything you assumed or guessed as UNCONFIRMED.

Start the block with: "Continue from this handoff. Ask me before acting on anything marked UNCONFIRMED."

The UNCONFIRMED label matters most. It catches the early guesses the study describes, and makes the new chat ask before building on them.

How to use it

1. Paste the handoff prompt into the old chat and send it.
2. Read the block before you copy it. Look hard at the UNCONFIRMED items, your rules and the numbers, and fix anything wrong right there in the text.
3. Open a new chat and paste the block as your first message. Re-attach any files it mentions by name.
4. Answer its questions about the UNCONFIRMED items, then carry on.

Run it when the answers start to slip (it forgets a rule, repeats itself, or brings back something you rejected), or when the work moves on to a new task.

Common mistakes

• Waiting until the chat is a mess. By then the early details may already be blurred. Run the handoff at the first sign of slipping, while the chat still has them straight.
• Pasting without reading. The block can carry a wrong number or a guess nobody marked. Read it before you paste, or the new chat starts from the same mistakes.
• Forgetting the files. The handoff names files and links without copying them. If you don't re-attach them, the new chat works from the names alone, or guesses.

The honest part

The handoff can only include what the old chat still has. In a very long chat, early details may already be summarised or garbled, so check names, numbers, dates and quotes against your originals before you rely on them. The AI also won't mark every guess. UNCONFIRMED catches the ones it notices, and some will slip through.

Think about what goes into the block. If the chat had account numbers, health details or other people's personal data, black them out before you paste it anywhere, and take extra care before pasting into a different tool or account.

Memory can bring old guesses back. If memory is on in ChatGPT or Claude, a new chat may still know things from the old one. If an answer leans on something that isn't in your handoff, ask where it came from.

This is for work: drafts, plans, research, code. If the long chat was about a legal, medical or money decision, or about how you've been feeling, a fresh chat won't fix that. Take it to a real person who can help.

Your action plan

1. Save the handoff prompt in a note you can find again.
2. Pick a long chat you're still working in and run the prompt there.
3. Read the block, fix what's wrong, and paste it into a new chat.
4. Ask the new chat: "What's the next step, and what are you unsure about?" If its answer matches yours, the handoff worked.

Start with the chat you have open right now.

Try next: Pick the right Claude model: Haiku, Sonnet, Opus or Fable