Paste This Into Claude Code, Never Run Out Of Tokens Again transcript

Claude just told me to come back in 5 hours and I still have a lot of work to do today. I didn't have it do anything massive. I didn't even run it for 8 hours straight. So, I set myself out on a journey to understand why I kept reaching my session limit and how I can stop that from happening ever again. [music] And the crazy part is almost none of my usage limit was because of what I typed. 0.01% [music] of it to be exact.

That is just how these tools work. And once you see the mechanism behind it, you can cut most of your costs down. So, in this video, I'm going to give you a prompt you can paste [music] into Claude code that audits your setup, then the seven fixes that actually move your token consumption [music] down. Starting with the one that costs nothing and ending with a trap that is probably [music] doubling your bill right now.

Let's get started. Here is the best way for me to explain token consumption. These models have no memory. None. Absolutely none. So, every time you hit enter, your entire conversation gets wrapped up and sent again from the top. Your first message costs what you typed. Your second message costs what you typed plus the answer plus the first message. By message 20, the thing you just wrote is that silver box at the top.

It's every single thing here. Everything under it is stuff you already paid for being paid for again. That is why it compounds the way that it does. A 3,000 token file your agent reads at turn four of a 40-turn session does not cost you 3,000 tokens. It costs you 3,000 tokens 37 more times. I mean, look at this messages line right here. That is not what you asked. That is all of this chat's history compounding over and over and over again.

And the reason that matters is that your setup is not the same as mine. So, let me show you how to find your own version of this. This is the prompt. I've included it in the description for free. You don't have to give me your email for it. Paste it into Claude code and it will audit your actual configuration. It reads your context breakdown, checks whether tool deferral is on, measures your memory files, looks at your cache hit ratio, and flags scheduled tasks that are firing while you sleep.

You can screenshot this video right now and send this to your AI agent. This will let you know what is consuming the most amount of tokens for you and which of these next seven fixes matter the most for you. Mine came back with these problems and I'm fixing them in order of what actually saves me the most. The highest leverage thing you can do to cut costs, costs you nothing and is five letters long. When you finish a job and start with a different one, use {slash}clear.

Do not keep going in the same thread because it is it is convenient. That old conversation is not sitting there quietly. It is being resent on every message you send until the session ends. Watch this. Messages are at 80,000 tokens and if I write {slash}clear and hit enter and boom, just like that, messages are back to 0%. And Anthropic's own docs say it plainly. When you want a fresh start instead of continuity, {slash}clear costs you nothing.

Here's why this beats everything else in this video. Every other fix I'm about to show you reduces one component of your context. Clearing resets the entire base that all of those components are a fraction of. 96% of my spend was rereading history. This is the one tool that deletes the history. One thing before you go clearing everything, use {slash} rename inside your session first, so you can use {slash} resume later, which allows you to restore the session in case you ever urgently needed it.

You are not throwing the work away, you're stopping the next job from carrying it. And that fix is free. The next one is the opposite, because it is something you are probably already doing on purpose to save money. This is the one I want you to actually remember. When you're running low, what do you do? You probably switch to a cheaper model. You hit {slash} model, drop from Opus to Sonnet, and you feel great about it.

That switch is the most expensive thing you can do. Here's why. Your conversation is cached, and cache reads cost 1/10 of normal output. That is why long sessions do not bankrupt you. But, the model is part of the cache key. Change the model, and none of your history matches the cache anymore. So, the entire conversation gets reprocessed at full price. On Opus 5, at 200,000 tokens of context, that turns a 10-cent turn into a 1-dollar turn. 10 times more expensive.

It's absolutely invisible, and you did it to yourself while trying to save money. Effort level does the same thing, by the way. Fast mode does the same thing, as well. So, things that break it, switching models, changing effort, turning on fast mode, connecting or disconnecting an MCP server if your tools load up front, enabling a plugin that ships an MCP server, and compacting. Also, and this one is nasty, upgrading cloud code and then resuming a long session.

Anthropic's docs literally call that the most expensive request you will send. Things that are safe, editing files in your repo, editing your memory file, changing output style, changing permission mode, invoking skills and commands, recaps, rewinds, and spawning a sub agent. So, the rule is simple. Pick your model and your effort at the start of the session and then leave those settings alone. If you want to be on a cheaper model, start there.

So, that is the expensive keystroke. The next three fixes, however, aren't something you've ever typed. Yet, they burn away at your context. Picture this, you type in install remotion for me. Your agent goes off and runs the actual commands and 800 lines come back. Package names, version numbers, warnings, a funding message. You did not read a single one of them. You just wanted to know that it actually worked and it installed the thing you asked it for.

But, your agent does not get to skim. All 800 lines went into your conversation and you pay for them again on every message until you clear. So, put a filter in front of it and you do not write it this next part. You asked for it. You can give it a prompt like the one I'm showing on the screen as I'm talking right now. It will create one small file that sits between your agent and the command and it cuts the output down before your agent ever sees it.

Your agent writes it, your agent installs it, and from then on it just works. Anthropic ships a working version of this so your agent has something to copy from. Their online on it is reducing context from tens of thousands of tokens to hundreds. You do this once and it works on every session after that. And this fix is output coming in. The next fix is what is already sitting in your context before you have ever typed anything.

You connect Gmail, then Notion, then Slack, and each takes you a single prompt to install, and it feels free, but it's not free. Every tool you connect comes with an with an instruction manual, what it can do, what to send, what comes back, and your agent has to read that manual before allowed to touch the tool. GitHub on its own, for example, costs you 26,000 tokens. Slack is 21,000 tokens. All of that gets loaded into every single session before you type a single word.

Now, here's the good news. Claude has released an update that makes sure your agent does not read every manual anymore. It loads the contents page, and it only opens the section it actually needs when it needs it. You still have access to the same tools, but that cost has went down 85%. That part is handled for you. It is on by default, and you do not have to do anything. But, you are still paying for the contents page.

And the contents page grows every single time you connect something new. So, here's the fix, and it takes about 30 seconds. Type in And just with that, you will see a panel of every single tool you have ever connected in one list with a switch next to each one. Go down in that list and turn off anything you have not used in the last like month. You're not deleting it, it stays set up, it just stops loading. One last thing to check while you are here, run /context, and find the tools line when it loads up.

So, we'll go ahead and expand, and there we go. We have system tools and system tools right here. If it says deferred, you are on the new behavior and the manuals are staying shut when you don't use them. So, as you can see, it's only using up 17,000 context tokens and 20,000 of the active tokens of skills I am currently using in this session right now. And the good news is turning tools off mid-session does not cost you anything.

As long as that line says deferred, connecting and disconnecting just appends. It does not rebuild your cache the way switching models does. And the next thing everybody tells you to do is delegate to sub-agents. So, I want to be honest about that because what you have been told is sort of half true. Everybody says sub-agents save tokens, but they do not. They move them. Here are Anthropic's own numbers from their documentation.

So, over here we have a simulation of how context windows work. And if I go ahead and go through the simulation, we get a prompt like use a sub-agent to research this session. And if you look at the context window, when we hit send, boom, just like that, the sub-agent has spent very few tokens because what we see is a sub-agent reads about 6,000 tokens of files and what comes back to your main context is a 420 token summary.

That looks like a massive win. And in your main window, it is. But, do the math on the whole thing. That sub-agent loaded its own system prompt, its own copy of your memory file, its own tools, and then did the reading. It burned roughly 9,800 tokens to save you 5,700. In isolation, you lost. And Anthropic is blunt about this elsewhere. Their multi-agent research post says agents use around four times more tokens than chat and multi-agent systems about 15 times more tokens than chat.

But, when is using sub-agents actually worth it? The answer is when three things are true. The output is high volume, you will not need the detail again, and the session is going to continue for many more turns. And that third one is the whole game because those 5,700 tokens you avoided would have been resent on every remaining turn. But, if you delegate and then immediately end the session, you just paid extra for nothing.

One free upgrade if you do this, set the sub-agents model to Haiku. That is a five times reduction on the isolated work and it does not touch your main session's cache. And that brings me to picking models properly, where there's a myth I want to kill. Half of what you ask for is small. Rename these files, write me a commit message, clean up this list. The useful heuristic is to use the dumbest model that will still finish the job.

But, remember fix two. Pick the right model at the start of the session. Switching mid-session costs you more than the saving. The better way to do this is per skill and per sub-agent. So, you can run Haiku for the grunt work without ever touching your main session model and without invalidating anything. Every fix so far assumes you are sitting at the keyboard. The next one is what happens when you're not. This is the trap I promised you at the start.

A scheduled task fires on its interval whenever you are there or not. And every time it fires, it sends your full context, not a bit of it, all of it. So, if that task is attached to a bloated session, you're paying for that entire context on every fire forever at 3:00 in the morning, while you're asleep. Now, here is the part that turns it from expensive to painful. Your cache expires. On a subscription, it lasts 1 hour.

So, if your task runs less often than once an hour, every single fire misses the cache and reprocesses your whole context at full price, instead of the 1/10 cache price. 10 times the cost on a schedule, forever. So, how often your task runs is a real cost setting. If your task can run every 45 minutes instead of every hour or 2 hours, it is cheaper to run it more often. And while I'm here, a correction on the opposite claim.

People say leaving Claude code open in the background burns your limits. Anthropic's documents background usage at under 4 cents a session. That is not your problem. Your scheduled tasks are your problem, and your live agent teams, because each one keeps consuming until it exits. That is the one that doubles your bill without you touching the keyboard even. And those are the seven that burn your token consumption number.

So, let me clear out the advice that does not, because some of it is actively wrong. All right, let's talk about what does not work. Number one, writing shorter prompts. In my logs, everything I actually type came to 0.01% of the bill. Your prompt length is a rounding error. Vague prompts do cost you, but through the file reads and the rework they trigger, not through length. Number two, compacting to save tokens. This one is backwards.

To write you a summary, it has to send your entire conversation one more time. So, the thing you did to save money is the single most expensive message of the session. And then it wipes your cache on purpose, because the conversation is summarized is no longer exists. Clearing is free. Compaction buys you continuity, not savings. And if you only want to undo a few back turns, use {slash} rewind instead. Rewind takes you back to a point your cash already knows, so nothing has to be reread.

Number three, screenshotting text to save tokens. A picture is not cheaper than the words in it. On Opus 5, one screenshot of your screen costs you about 2,700 tokens. A 4K one is nearly 5,000. That is a lot of text. Paste the text instead. It is cheaper and your agent can actually edit it. It cannot edit a picture though. And finally, PDFs, while I'm on format. Every page of a PDF costs you between 1,500 and 3,000 tokens just for the words on it.

And then your agent also takes a picture of that page and you pay for that picture, too. You basically pay for every page twice. Ask your agent to turn it into a plain text file first and the same document costs you about a quarter as much. And there is one last thing you need if any of this is going to stick. Four things, three of them are already sitting in your terminal and you have probably never opened them. {slash} context.

This is the one you're going to use the most. It shows you what is in your window right now, line by line, so you can see exactly what is taking up the room. Then we have {slash} usage. This one is underrated. It shows you how much of your plan you have burned through. And then it tells you what burned it, not roughly. It names the specific skill, the specific tool, the specific agent. So if one thing on your machine is quietly eating your limits, this is where it confesses.

Then inside of that, you have what is this one session costing you? The cost or slash cost. And how much of it was your agent re-reading history versus doing new work. And the fourth one is the burn rate meter that is in the corner of your cloth code screen. Having a number moving while you work changes your behavior more than any rule I have given you. And if you want the number I opened this video with, it is sitting on your machine right now.

Every session you have ever run is logged in a folder. And every single reply in there records what it costs. So, ask your sub agent to go read them and work out your own percentage. That is literally what I did. And then, you have your number instead of mine. You did not run out of tokens because you asked too much. You ran out because almost everything you paid for was your agent re-reading things you already sent, and nobody ever showed you where the switch was.

So, now you know. Clear between jobs because it is free and it resets the base everything else is a fraction of. Pick your model and effort once and leave them alone because the thing you do when you are trying to save money is the most expensive move on the board. Filter your tool output before it lands. Disconnect the tools you never use and check that line in context stays deferred. Delegate when the session has a long way left to run and not when it does not.

And go look at what your scheduled tasks are doing at 3:00 in the morning. One honest note to finish on, the labs are not going to fix this for you. Not because they're being difficult, but because they're not graded on how few tokens you use. It is your desk. You have to keep it clean. The token usage audit prompt is in the description and I would run it once a week or whenever you feel you've been burning through a lot of your tokens fast.

Things drift, you add a server, you install a plugin, you change a setting, and 6 weeks later you are back where you started wondering why you are hitting limits again. Comment below with what shocked you the most in this video. And if you enjoyed this video, make sure to leave a like. And if you're new to my channel, then subscribe because I have a ton more content like this coming your way. And oh, what would you know?

Over here, the algorithm gods seem to think that you will really enjoy this video. So, click it, and I'll see you there.