Skip to main content

10. AI Tokens - Management

Note: The video covers material not in the guide below — please watch in full.

Action Step

Complete this before moving on.

Follow along with the video in your own repo. Create a handoff document by telling Claude to save a detailed summary of your current session into a file before compacting. Then try the chain-linking workflow: create a scoping document with a to-do list for a project, and have Claude start working through the phases. As you get more repetition with this, you'll start to feel when it's time to save your work and start fresh.

Comment in Slack

Post your answer in your onboarding channel.

What was your biggest takeaway(s) from this training?

Training Guide

You know what tokens are. You know the context window has a limit. You've seen what happens when compacting kicks in — the telephone game.

This training is about what to DO about it. How to manage your context so you never lose your work, and how to scale beyond any single conversation.

(Let's start with your options when the warning hits)

Three Ways to Handle It​

When you see the token warning — and you will, eventually — you have three options:

Option 1: Let it auto-compact. Do nothing. When you hit 100%, the AI compacts on its own and picks what to keep. This is Russian roulette with your context. Sometimes it's fine. Sometimes it drops the one thing you needed it to remember.

Option 2: Manually compact early. When you see the warning, you hit compact yourself. Better than auto — you're choosing when it happens instead of letting it surprise you. But it's still a summary, and details still get lost.

Option 3: Save your work first, then start fresh. This is the move. Before you hit the wall, you tell the AI to write down everything important — what you've done, what's left, key decisions, important context — into a file. Then you compact or start a new conversation. The AI reads that file and picks up exactly where you left off.

Option 3 is the only one where you don't lose anything. The other two are gambling.

(Let's talk about how to actually do Option 3)

The Handoff Document​

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

Here's the concept: anything written to a file lives outside the conversation. It doesn't get summarized. It doesn't drift. It doesn't disappear when you compact. It's just a file on your computer that any future conversation can read.

When you're deep in a project and the token budget is getting low, you tell the AI: "Write a handoff document." It creates a file that captures:

  • What the project is and the goal
  • Key decisions made so far
  • What's done and what's next
  • Important context that can't be lost
  • Specific instructions for continuing

Then you start a fresh conversation, point it at that file, and say: "Read this and continue." The new conversation has a full 200,000 tokens to work with — and it knows everything the last one knew.

Guided demo: try it right now.

Tell Claude:

"Create a handoff document and save it as handoff.md in my folder. Document what we've been working on, what I've learned so far in these trainings, any important context, and what I should focus on next. Write it so that if I start a completely fresh conversation, the AI can read this and know exactly where I left off."

Look at what it creates. That's your template. From now on, before any long project hits the token wall — create one of these.

(There's actually a second type of document worth saving)

The Learning Document​

A handoff document is about continuing work. A learning document is about capturing what you discovered.

Think of it this way: you just spent an hour having an incredible conversation with the AI. You learned a ton. You had a few "OMG" moments. Some great tables got created. Some insights clicked.

Now close that conversation. Come back next week. Try to find that one insight from last Tuesday by scrolling through old chat history. Good luck.

A learning document solves this. Before you close any session where you learned something useful, tell the AI:

"Capture everything I learned in this session into a learnings.md file. Include key concepts, any tables we created, insights that clicked, and things I'd want to reference later. Keep it clean and scannable."

It takes 30 seconds. It saves you from ever digging through old conversations again.

(Now let's connect this to how you'll actually work on big projects)

Chain-Linking: How You Scale Beyond One Conversation​

Here's where context management goes from defensive to strategic.

Say you have a big project — a 15-page strategy doc with research, analysis, and recommendations. That's way more than one conversation can handle. So you break it into phases:

  1. Create a scoping document with a to-do list — every phase of the project, laid out as checkboxes
  2. Conversation 1 tackles the first few tasks, checks them off, and creates a handoff document
  3. Conversation 2 reads the scoping doc and the handoff, picks up where Conversation 1 left off, does the next set of tasks
  4. Repeat until the project is done

Each conversation gets a fresh 200,000 token budget. The scoping doc is the thread that ties them all together. The handoff documents make sure nothing gets lost between sessions.

This is called chain-linking — and it's how you run projects that would be impossible in a single conversation. The token limit stops being a wall and starts being a natural breakpoint.

One conversation can do a lot. Chain-linked conversations can do anything.

(And remember — the AI can also spin up copies of itself to work in parallel. That's sub-agents, and that's next)