Skip to main content

Managing 200k Token Window

Outcome: You'll handle projects larger than one context window without losing critical information.

The Token Limit Reality​

You have 200,000 tokens per session. Sounds like a lot—and it is—but complex tasks eat through it fast.

What consumes tokens:

  • Long source documents (a 50-page transcript = ~15,000 tokens)
  • Detailed prompts
  • Claude's responses
  • The entire conversation history

When you're running low:

  • You'll see a warning at ~50% usage
  • The "Compact" button appears
  • Claude may start "forgetting" earlier context

The Telephone Game Problem​

Remember the telephone game as a kid? You whisper a message down a line of people, and by the end it's completely different from how it started.

The same thing can happen with AI context. Every time you compact, Claude summarizes the conversation. That summary becomes the new "memory." If you compact again, it summarizes the summary. Details get lost. Nuance disappears. Instructions drift.

This is why token management matters. It's not just about fitting more in—it's about preserving the accuracy of your context over long projects.

The strategies below help you avoid the telephone game.

The Compact Feature: Your Choice​

Here's the reality: if you hit 100% of your token window, Claude will auto-compact. It's going to summarize whether you like it or not.

So you have a choice:

OptionWhat HappensThe Risk
Let it auto-compactClaude decides what to summarizeYou're playing Russian roulette with your context. Important details might get lost.
Proactively save your workYou create a handoff document before hitting the limitTakes a few minutes, but you control what's preserved.
Manually compact earlyYou click Compact when you see the warningBetter than auto, but still a summary of a summary over time.

The best practice: Don't wait for auto-compact. When you're deep in a project and see the warning, take a moment to save your key insights and context to an external file. Then you can compact (or start fresh) without losing what matters.

Using External Files to Preserve Continuity​

The handoff document is your insurance policy against the telephone game.

The concept: Anything written to an external file lives outside Claude's token window. It doesn't get summarized. It doesn't drift. When you start a fresh session, Claude can read that file and pick up exactly where you left off.

What to include in a handoff document:

  • What the project is and the goal
  • Key decisions made so far
  • Current status (what's done, what's next)
  • Important context that must not be lost
  • Any specific instructions for continuing

Hands-On: Create a Handoff Document​

Let's practice this now so it becomes second nature.

  1. Create a new file: handoff.md in your test folder
  2. Copy the path
  3. Give Claude this prompt:
Create a handoff document and working log in this file: [paste path]

I want you to document:
- What we've been working on in this session
- Key things I've learned (for example: what tokens are, what the agent platform features are, how prompting works)
- Any important context that should carry forward
- What I should focus on next

so that before you compact, we can have as much context saved and you know where to pick up from where we left off.

Write this so that if I start a completely fresh session, I can read this document and know exactly where I left off.
  1. Review what Claude creates
  2. This is now your template for future projects

From now on: Before you hit token limits on any big project, create a handoff document. Future you will thank you.

Hands-On: Create a Learning Document​

Different from a handoff doc, a learning document captures what you discovered—insights, findings, interesting things that came up.

Why bother? Yes, you can technically go back and look at historical conversations in Claude. But it's annoying. Scrolling through old chats trying to find that one insight from last Tuesday? Not fun. Better to just grab the good stuff before you close.

  1. Create a new file: learnings.md in your test folder
  2. Copy the path
  3. Give Claude this prompt:
I want to capture all of my learnings into this file: [paste path]

Document the interesting insights from our conversation. Include:
- Key concepts I learned
- Any tables we created (I really liked those)
- Direct explanations that clicked for me
- Things I'd want to remember later

Keep it clean and scannable.
  1. Review what Claude creates
  2. Now you have a reference doc you can actually use

The habit: Before closing any session where you learned something useful, ask Claude to dump your learnings into a file. Takes 30 seconds, saves you from digging through chat history later.

Multi-Agent Preview​

One more tool for handling large projects: Claude can spawn sub-agents.

Each sub-agent gets its own fresh 200,000 token window. They work in parallel and report back. Instead of one agent struggling with a massive task, you have a team.

Example: You need to research 10 competitors. Instead of doing them one-by-one (eating up your tokens), Claude spawns sub-agents that each handle one competitor simultaneously.

We'll go deep on this in Section 3.2. For now, just know it's another way to scale beyond the single token window.