Does the compact command fix a long Claude Code chat?

It shortens the chat, but it replaces it with a summary of the entire conversation. The dead ends and bad guesses are summarized in, so they stay.

What is a handoff file?

A short note, usually called handoff.md, that Claude writes at the end of a session. It lists the goal, the current state, the files involved, the changes made, what failed and the next steps. A fresh session reads it to pick up where the old one stopped.

How do I get a /handoff command in Claude Code?

Make a skill. Create a folder called handoff inside .claude/skills with a SKILL.md file that holds your handoff prompt. The folder name becomes the command, so you can run /handoff.

It is the name Chroma uses for AI models getting less reliable as the input gets longer. Chroma tested 18 models and found their performance grows increasingly unreliable as the input gets longer.

Skills

Keep Claude Sharp in Long Sessions With a Handoff Skill

5 minute readUpdated October 2026Explore more

TL;DR

The longer a Claude Code chat runs, the worse the answers get. Chroma tested 18 AI models and found they got less reliable as the input grew. Compact does not clean that up, because it keeps a summary of the whole chat, dead ends included. The fix is a handoff file: ask Claude to write down what is done, what failed and what is next, then start a fresh session that reads it. Save that prompt as a skill called handoff and the whole move is two words: /handoff.

Learn Claude Code. Earn income. Only $9.

👉 https://www.skool.com/claudecodeclub

Why a long Claude Code chat gets worse

Claude reads the whole chat every time it answers. When the chat gets long, the answers get worse. Chroma tested 18 AI models and found their performance grows increasingly unreliable as the input gets longer. Chroma calls it context rot.

The Chroma technical report called Context Rot: How Increasing Input Tokens Impacts LLM Performance — Chroma's Context Rot report, July 2025

Why compact does not fix it

The compact command shrinks your chat to save space. Claude Code's docs say it replaces the conversation with a structured summary of the entire conversation. So the wrong turns you took an hour ago come along inside the summary. You get a shorter chat, but it is the same chat.

Step 1: Ask Claude to write a handoff file

A handoff file is a short note from your current session to your next one. It holds only what the next session needs, with none of the back and forth. Ask for it before the chat gets bad, or the moment you notice Claude going in circles.

promptWrite a file called handoff.md for this session. Include: the goal, the current state, the files we are working on, the changes we made, what we tried that failed, and the next steps. Keep it short and specific.

Step 2: Start a fresh session and read it

Open a new Claude Code session in the same folder. The new session starts with an empty chat, so nothing from the old one weighs it down. Then point it at the file.

promptRead handoff.md, then tell me in two sentences where we left off and what you will do next. Wait for my go before you change anything.

Step 3: Save the handoff prompt as a /handoff skill

Claude Code's docs say to make a skill when you keep pasting the same instructions into chat. A skill is a file named SKILL.md inside a folder under .claude/skills. The folder name becomes the slash command, so a folder called handoff gives you /handoff.

The Claude Code docs page called Extend Claude with skills — Claude Code docs: skills

You do not need to write the file by hand. Ask Claude to build it:

promptCreate a skill called handoff at .claude/skills/handoff/SKILL.md. When I run /handoff, it should write handoff.md with: the goal, the current state, the files we are working on, the changes we made, what failed, and the next steps. Keep it short and specific.

From then on, ending a session takes one word: /handoff. Open a new session, say read handoff.md, and keep going.

Best practices for handoff files

Write the handoff early. Do it when Claude starts repeating itself or fixing the same bug twice, not after the chat is a mess.
Always list what failed. The next session will try the same dead ends unless the file says not to.
Keep it short. A page of facts beats ten pages of story. The new session should read it in seconds.
Check it before you start fresh. Open handoff.md and fix anything wrong. A bad handoff carries a bad guess into the new session.

Which move to start with

Claude is looping right now: run the Step 1 prompt, then do Step 2 in a new session.
You do this more than once a week: build the /handoff skill in Step 3 so it is one word.
Your chat is short and going well: skip all of this. Handoffs are for long sessions.

Learn Claude Code. Earn income. Only $9.

👉 https://www.skool.com/claudecodeclub

Common questions

Does the compact command fix a long Claude Code chat?
It shortens the chat, but it replaces it with a summary of the entire conversation. The dead ends and bad guesses are summarized in, so they stay.
What is a handoff file?
A short note, usually called handoff.md, that Claude writes at the end of a session. It lists the goal, the current state, the files involved, the changes made, what failed and the next steps. A fresh session reads it to pick up where the old one stopped.
How do I get a /handoff command in Claude Code?
Make a skill. Create a folder called handoff inside .claude/skills with a SKILL.md file that holds your handoff prompt. The folder name becomes the command, so you can run /handoff.
What is context rot?
It is the name Chroma uses for AI models getting less reliable as the input gets longer. Chroma tested 18 models and found their performance grows increasingly unreliable as the input gets longer.

Keep going

Want the skills that make Claude Code worth every token?

Get 650+ plug-and-play skills, MCPs & prompts, plus 8,000+ members - $9/mo, cancel anytime.

Join the Club