May 21, 2026 · Free Guide

Anthropic Says Start a Fresh Claude Code Session Every Task (Here's the One-Command Handoff That Saves Your Place)

You ever notice Claude Code starts repeating itself around message 25? Forgets what file you were editing. Suggests the same fix you already tried. That's not you. That's context rot, and Anthropic just publicly said it's a real thing.

Why Long Claude Sessions Get Worse, Not Better

Direct quote from the Anthropic team in their official session management post:

"Context rot is the observation that model performance degrades as context grows because attention gets spread across more tokens."

Translation. Every message you send adds tokens to the conversation. The longer the chat, the more the model has to scan. Old irrelevant stuff competes with the thing you actually care about right now.

Performance drops. Not in a small way.

THE OFFICIAL RULE

It's Not About Message Count. It's About Tasks.

Most people online say "start a new conversation every 20-30 messages." That's a community heuristic. Decent rule of thumb. Not what Anthropic actually said.

The real recommendation, quoted directly from their post:

"When you start a new task, you should also start a new session."

Task-based. Not message-count-based. Finished debugging that build error? New session. Done researching, ready to write? New session. Switching from refactoring to writing docs? New session.

The Catch Nobody Mentions

Starting fresh sounds great until you realize you just lost 30 messages of context. The files you touched. The dead ends you already ruled out. The exact error you were chasing. The decision you made on line 42.

You either spend 5 minutes typing out a summary, or you start fresh and Claude tries everything you already tried.

Anthropic actually addressed this in the same post. They recommend using /rewindor "summarize from here" to:

"Create a handoff message, kind of like a message to the previous iteration of Claude from its future self."

Past Claude leaves notes for future Claude. Good idea. The execution is rough though, because the output is loose and unstructured.

THE SKILL

/handoff. One Command. Structured Output. Copies Itself.

I built a free Claude Code skill that handles this in one command. You type /handoff, it reads your whole session, distills it into a structured message, and copies it to your clipboard. Paste into a new session and pick up exactly where you left off.

What the output includes:

  • What you're trying to do (the goal)
  • Where you left off (current state)
  • Key files and paths (absolute, ready to Read)
  • What was about to happen next
  • Things that already failed (so the next Claude doesn't repeat them)
  • Important context (IDs, URLs, decisions)
  • The first concrete action for the new session

Stays under 400 words. No tool-call noise. No restating CLAUDE.md (the next session auto-loads that anyway).

Install It in 30 Seconds

Open Claude Code and paste this message:

Install this skill for me: https://github.com/tenfoldmarc/handoff-skill

Claude will clone the repo into the right folder. After that, any time your session is getting long, type /handoff, paste the output into a fresh window, and pick up where you left off.

When to Use /handoff vs /clear vs /compact

/compact

Built into Claude Code. Summarizes the conversation and replaces the history with the summary. Same session, just shorter. Use when you want to keep going in this window.

/clear

Wipes the conversation. Use when you're switching to an unrelated task and don't need any of the previous context.

/handoff

Use when you want to start fresh in a NEW window but keep your place. Best for the "I'm at the natural end of this work but I'll come back to it later or in another terminal" case.

Rule of thumb: stuck in the same session? /compact. Done forever? /clear. Coming back fresh? /handoff.

All Resources

Anthropic: Session Management & 1M ContextThe official blog post quoted throughout this guide
/handoff skill on GitHubFree. MIT licensed. One git clone to install.
Claude Code Best PracticesOfficial Claude Code docs from Anthropic
Claude Code Compaction DocsHow /compact works under the hood

Prefer watching over reading?

Watch the Reel on Instagram →

The Next Step

Guides show you what. AI Builders shows you how.

Inside the community, I walk through every build live — including the stuff that doesn't make it into guides. Regular people (not developers) figuring out AI together, shipping real projects, asking me anything. No fluff, no theory, just the actual work.

Join AI Builders

skool.com/ten-fold