Never run out of context again

If you’re anything like me, you’ve racked up a 1,000,000-token session before realizing what happened.

You start a chat.

It’s going great. The AI remembers everything, follows your logic, catches its own mistakes.

Then, somewhere down the line, it changes.

It re-asks things you already answered. It contradicts a decision you made 10 messages ago. It gets slower. Dumber…

That’s context rot.

The old fix? Scroll up. Re-explain everything. Copy-paste your last few messages and hope the AI catches up.

There’s a better way: the handoff.

You ask your AI to summarize the whole session into one clean document, right before you leave the chat. Then you paste that document into a brand new chat, and the new chat picks up exactly where the old one left off, minus all the dead weight.

If you’ve used any AI chat before, you can start doing this TODAY.

What you need before you start:

• Any AI chat tool. Claude, ChatGPT, Gemini, whatever you’re already using.
• A project worth protecting. Something with more than a handful of messages, decisions, or working code.
• Nothing else. No install, no plan upgrade required for the basic version.

First, why does this even happen?

Every message you send and every reply you get sits inside something called a context window. Think of it as the AI’s short-term memory for that chat.

The longer the chat runs, the more that window fills up. Old details start getting pushed out or blurred together to make room for new ones. That’s context rot: the AI is still “there,” but its grip on the earlier details is loosening.

You’ll notice it as:

• Repeated questions about things you already settled
• Decisions quietly getting undone or ignored
• Slower, vaguer responses
• The AI reaching for a fix you already ruled out 15 messages ago

None of that means the AI is broken. It means the chat is full.

You don’t fix a full suitcase by shoving harder. You repack it.

Here’s the actual fix.

Once you’re around 20 messages deep, or you feel the AI starting to slip, ask it to write a handoff document. This is a structured summary built for a brand new AI instance that knows nothing about your project yet.

Here’s the prompt:

Create a handoff document for a new agent instance with zero context on [PROJECT NAME] with the below:

1. What this is + current state — architecture, what’s built, what’s tested vs not

2. Key decisions made and why (so they don’t get re-litigated)

3. Constraints — explicit requirements or things ruled out

4. Errors/roadblocks hit and their root cause — enough that the next agent doesn’t burn context or credits repeating dead ends

5. Immediate next steps, ordered

Swap in your actual project name, and let it run.

You’ll get back a tight, structured document. Not a transcript, not a recap of every back-and-forth. Just the parts that matter going forward.

Then open a brand new chat, and paste it in as your first message.

The new chat starts with all the important context and none of the clutter. It’s faster, sharper, and won’t repeat mistakes the old chat already burned tokens learning.

Now that you’ve got the basic handoff down, here’s the next level.

If you’re building something with a lot of moving parts, don’t try to do it all in one marathon chat. Split it by function instead.

One chat for the backend logic. A separate chat for the UI. Another for testing. Each one gets its own handoff document when it wraps up or hands off to the next stage.

Why bother? Because context rot doesn’t just cost you tokens. It costs you accuracy. A chat juggling five different concerns at once is more likely to blur them together than a chat with one clear job.

Keep each chat scoped to one function, and hand off between them the same way: summarize, paste, continue.

Here’s the part that actually saves you.

Remembering to ask for a handoff at the right moment is its own chore. So don’t remember it. Offload it.

Most AI chat tools let you set custom instructions, standing directions the AI follows automatically in every conversation. Add one that tells it to watch your context usage and speak up.

Here’s the instruction to add:

When this conversation’s context usage crosses [YOUR THRESHOLD], stop and tell me. Suggest that we generate a handoff document before continuing, using the 5-part handoff structure (current state, key decisions, constraints, errors/roadblocks, next steps).

Now the AI watches the clock for you. You just keep working until it taps you on the shoulder (test that it works first ;)).

And if you’re lazy like me, here’s the last step: stop copy pasting it over and over again.

Instead, turn it into a reusable skill. Most AI platforms let you save a prompt once and summon it with a short trigger, a slash command…

Set it up once. From then on, you just type something like /handoff, and the full 5-part structure runs without you retyping a word.

See my guide for creating a skill:

Once this is set up, long projects stop being a liability. You can run a chat as long as you need, hand it off clean, and keep going without losing the thread.

No re-explaining.

No re-litigating old decisions.

No 1,000,000-token panic.

A long project isn’t one long chat. It’s a relay, and the handoff document is the baton.