Free skill
Save your chat. Start the next one in seconds.
- Saves what you decided, what is done and what is next into one file.
- Writes the prompt that starts your next chat exactly there.
- Long chats cost more every message. A fresh one does not.
1. Copy it
Close this chat so the next one starts where this one stops, without carrying the whole history. Do three things, in this order. Be specific. Use the user's own words wherever they decided something. 1. Write the HANDOFF. Keep it under 500 words. Sections, in this order, skip a section only if it is empty: - Goal: one sentence. - Decisions: each one the user made, quoted word for word, with the date if known and the reason in a few words. - Facts and numbers: each with where it came from (file, command, link). Anything with no source is marked UNVERIFIED. - Files and links: every file this chat created or changed, with its full path or link. - Rejected: what was tried or refused, and why. This stops the next chat from suggesting it again. - Unfinished: what is not done, and the exact next step. - How the user likes to work: language, tone, length, anything they corrected. Quote what you can see; if a message is cut off, quote the visible part and mark it (cut). A claim from a helper or tool is a claim, not a fact: name who made it. If you must cut for length, cut file detail first and keep decisions, the rejected list and the next step. Never invent. If you do not know, write UNKNOWN. Do not paste raw tool output or long file contents; name the file instead. 2. Save it. - If you can write files (Claude Code): write it to HANDOFF.md in the current project folder. If the file exists, replace its "Current state" section and keep the rest. - If you cannot (claude.ai): show it as one block the user can copy, and say: "Save this in your Project's knowledge, or paste it as the first message of the next chat." 3. Write the start prompt for the next chat, in one code block the user can copy. It must work with nothing else in context. Format: - One line with the goal. - "Read HANDOFF.md first" with its full path (Claude Code) or the HANDOFF text itself (claude.ai). - "Do not re-open the decisions listed. Start with: <the exact next step>." - "Ask me before doing anything outside that step." Then tell the user in one line: the new chat is ready, they can close this one.
2. Install it (60 seconds)
On claude.ai
- Open a long chat you want to leave.
- Paste what you copied as your next message and send.
- Claude shows the handoff and a start prompt. Paste the start prompt into a new chat.
To make it permanent: open a Project, add the text to Project instructions, then type "handoff" in any chat there.
In Claude Code
- Tap Download. You get
SKILL.md. - Put it in
~/.claude/skills/chat-handoff/SKILL.md. Or paste the next block into Claude Code and let it save the file for you. - In any chat, type
/chat-handoff.
Or let Claude Code save it
Save the text between the markers as ~/.claude/skills/chat-handoff/SKILL.md, exactly, then confirm the path. ----- --- name: chat-handoff description: Use when the user says handoff, wrap up, save this chat, or next chat, or when the conversation has grown long. Writes the chat's state to HANDOFF.md and a start prompt for the next chat. --- Close this chat so the next one starts where this one stops, without carrying the whole history. Do three things, in this order. Be specific. Use the user's own words wherever they decided something. 1. Write the HANDOFF. Keep it under 500 words. Sections, in this order, skip a section only if it is empty: - Goal: one sentence. - Decisions: each one the user made, quoted word for word, with the date if known and the reason in a few words. - Facts and numbers: each with where it came from (file, command, link). Anything with no source is marked UNVERIFIED. - Files and links: every file this chat created or changed, with its full path or link. - Rejected: what was tried or refused, and why. This stops the next chat from suggesting it again. - Unfinished: what is not done, and the exact next step. - How the user likes to work: language, tone, length, anything they corrected. Quote what you can see; if a message is cut off, quote the visible part and mark it (cut). A claim from a helper or tool is a claim, not a fact: name who made it. If you must cut for length, cut file detail first and keep decisions, the rejected list and the next step. Never invent. If you do not know, write UNKNOWN. Do not paste raw tool output or long file contents; name the file instead. 2. Save it. - If you can write files (Claude Code): write it to HANDOFF.md in the current project folder. If the file exists, replace its "Current state" section and keep the rest. - If you cannot (claude.ai): show it as one block the user can copy, and say: "Save this in your Project's knowledge, or paste it as the first message of the next chat." 3. Write the start prompt for the next chat, in one code block the user can copy. It must work with nothing else in context. Format: - One line with the goal. - "Read HANDOFF.md first" with its full path (Claude Code) or the HANDOFF text itself (claude.ai). - "Do not re-open the decisions listed. Start with: <the exact next step>." - "Ask me before doing anything outside that step." Then tell the user in one line: the new chat is ready, they can close this one. -----
What it looks like
You get a HANDOFF with your goal, your decisions in your own words, the files, what was rejected and the exact next step. Then one prompt for the next chat. Tested on a real chat of 327,104 tokens: the next chat started at about 71,800 (4.6x less), and the handoff itself was about 1,000 tokens.
Free to use and share. It is plain instructions, nothing is installed on your computer.