Claude Code: How to Cut Your Token Spend

Today I want to show you how to cut your token spend several times over. Not by a few percent — several times over.

The problem is the same for everyone. You work with an AI for half a day, and your tokens run out. Now you either wait until tomorrow, or pay more, or drop down to a weaker model that gets things wrong twice as often. Work stops. And it stops at the worst possible hour, right when the task finally started moving.

That is only half of it. By the evening the model also starts getting dumber. It forgets what you agreed on. It brings back the approach you rejected. It opens files it went through an hour ago.

You think: long day, lots done, the machine is tired.

Not even close. You spent the whole day paying for the same old garbage, over and over, on every single message. And that garbage is what made it dumber.

This happens to everyone. But almost nobody talks about it, and the usual advice is small change: write shorter prompts, ask fewer questions. That does not work.

Something else works. And almost nobody realises how it actually operates under the hood.

Let me show you with a plain example, and you will never work the old way again.

Picture a chat on your phone. An ordinary conversation. You send a message, the other person replies. You send a second one, they reply again.

Now picture that this person works in a strange way. Before every single reply, they reread the entire conversation from the very first message. All of it. Every time.

You have sent them forty messages. On the forty-first, they read all forty again. Start to finish.

That is exactly how an AI model works. Exactly.

It feels like you are paying for the message you just typed. You are not. You are paying for the whole conversation. Again. On every reply.

That whole thing has a name. It is called the context. The context is everything the model holds in front of its eyes when it writes your next answer: your messages, its answers, the files you pasted, the errors you dropped in. All of it, from the first line down.

A token is a chunk of text, roughly half a word. The model does not count your questions, it counts text. So here is the thing: it is not your question that eats your tokens. It is the context that grew around your question.

Now watch what follows from that.

Early in the session you pasted a big file. Or a two hundred line error log. Or a listing of a folder. You looked at it once and moved on.

The model did not move on. It reads that file again and again, on every message you send. Hour after hour.

You solved that problem a long time ago. The money for it keeps going out.

So garbage at the start of a conversation is far more expensive than garbage at the end. The very same file, pasted at the end, costs you almost nothing. Pasted at the start, it costs you a whole day.

And the worst part: nothing warns you. There is no alert. The model answers like normal. Everything looks fine. And then one day you hit the wall and think, well, I worked a lot today.

You did not work a lot. You dragged a dead weight around all day.

So what do you do? One command. Slash clear.

In Claude Code you type a forward slash and the word clear, and the whole conversation is wiped. Other tools call it something else: new chat, start over. It is the same thing.

It is free. And it will save you more than any clever prompt trick ever will.

The rule is short. Finished a topic — slash clear. Right now, not later.

Fixed the bug — slash clear. Answered the question — slash clear. Closed the task — slash clear.

Do not wait until the conversation gets heavy. By the time it feels heavy, you have already been charged for it.

Here people usually ask: if I run slash clear, do I not lose everything we agreed on?

You do. If you kept it in the chat.

And that is the second habit, and it matters even more than the first. Write it to a file.

I keep a plain text file. Before I hit slash clear, I write down what is done, what is next, what broke. Three minutes.

After that — slash clear, with no regrets. The new chat reads that file and knows everything it needs in ten seconds.

Here is the difference. For a chat you pay every single time. For a file you pay once, the moment it is loaded.

A chat is like a subscription that charges you daily. A file is like buying the thing once and owning it.

There is a third piece, and it moves the numbers more than anything else.

Do not keep one assistant running all day. Do it like this: one task, one short chat. Task done, chat closed for good.

Let us count it on a made-up example, one you can do in your head.

Eight similar tasks in a day. First way: one long chat for all eight. Feels convenient. In practice it is a disaster. By the eighth task it is carrying the entire history of the first seven. And you pay for that history again and again.

Do the maths. Task one is one portion of text. Task two is two: its own plus the first. Task three is three. And so on to the eighth, which is eight portions. Add it up: thirty six portions across the day.

Second way: one task, one chat. The assistant gets the job, does it, writes one line into a table and shuts down. Nothing carries over. Eight tasks, eight portions.

Thirty six against eight. Same work, same quality. A difference of four and a half times, and it comes out of thin air.

The model did not get smarter. You simply stopped paying for what nobody needs.

And you know what turned out to matter even more than the money?

A bloated context makes the model dumber.

Look at what you are asking it to do. You hand it a huge conversation where almost nothing still matters, and at the end you add a request. That request now competes with a file you pasted an hour ago and a bug you fixed long ago.

It does not know which parts are dead. As far as it can tell, everything you gave it was given for a reason.

So it starts opening files it already opened an hour ago. Bringing back a solution you rejected. Remembering a rule that stopped applying ages ago.

People say the model gets dumber by the end of the day. It does not. You buried your request under a pile of old junk.

Slash clear, and it is sharp again. You simply stopped making it guess which of a thousand messages still counts.

How do you know it is time to hit slash clear? Three signs.

One. You finished the topic. It works, the question is closed. Slash clear.

Two. You are explaining something for the second time. That means your request got buried. Slash clear, and say it once, at the top.

Three. You came in with one thing and now you are discussing three different ones. Three topics in one chat means each of them is paying for the other two.

Let me pull it together. Three habits.

First: finished a topic, slash clear, straight away.

Second: keep what matters in a file, not in the context. If your work does not survive a reset, that means you wrote nothing down.

Third: one task, one short chat. Done and closed.

This is not about being cheap. It is about where your money goes. Every token spent rereading old garbage is a token not spent on your actual task.

Slash clear early. Write to a file. Keep the context short.

That is all for today. My name is Evgenii Arsentev, and I bring artificial intelligence into companies: automating routine work and cutting operating costs. Thanks for listening.