18 Claude Code Token Hacks in 18 Minutes — Transcript
Full transcript
- 0:00In the past week or so, so many people
- 0:01have been complaining about hitting
- 0:02their Claude code limit insanely fast.
- 0:05Claims like one prompt that is about 1%
- 0:06of the limit is now around 10%. You
- 0:08could go through X and find tons and
- 0:10tons of threads about this topic. Even
- 0:12on a $200 per month plan, people are
- 0:14reaching the session limit way too fast.
- 0:17And then we got this post from an
- 0:18Anthropic employee that basically said
- 0:19that they are working on a little change
- 0:22with peak hours and off-peak hours. But
- 0:24even after that, some people were saying
- 0:26they were still hitting it really quick
- 0:27even during off-peak hours. So anyways,
- 0:29I've been playing around a ton, trying
- 0:30different doing research, and I have 18
- 0:33token management hacks for you guys that
- 0:34I've organized from tier one all the way
- 0:36up to tier three, so they get more
- 0:38advanced as we go. I'm very confident
- 0:40that by the end of this video, you will
- 0:41feel like your Claude code usage has
- 0:43doubled, tripled, maybe even five-xed.
- 0:45So let's not waste any time and just get
- 0:46straight into the video. So as I've been
- 0:48optimizing my own token management, I
- 0:50think that what's really important to
- 0:52realize first is how tokens actually
- 0:54work. Because once you realize how
- 0:56Claude uses tokens, it makes it very
- 0:58clear how you should actually reverse
- 1:00engineer the way that you work in order
- 1:02to use less tokens. So a token is the
- 1:04smallest unit of text that an AI model
- 1:06reads and charges you for. It's roughly
- 1:08one token is one word, but that's not
- 1:10explicitly true, kind of just a good
- 1:12baseline. So every time that you send a
- 1:13message, Claude rereads the entire
- 1:15conversation from the beginning, and all
- 1:17of those are tokens that it's charging
- 1:19you for. So message one, it will read
- 1:21it, then it will read its reply, and
- 1:22then message two, and then the reply,
- 1:24all the way up to your latest prompt.
- 1:26And it does that every single time. And
- 1:28I think that alone is a huge lightbulb
- 1:30moment for a lot of people. This means
- 1:32as you're having a conversation with
- 1:33Claude, your cost is compounding, not
- 1:35just adding. It's exponentially growing.
- 1:38Meaning message one might cost 500
- 1:40tokens, message 30 costs 15,000 because
- 1:43it's rereading everything before it. One
- 1:45developer actually tracked a 100-plus
- 1:47message chat and found that 98.5% of all
- 1:50the tokens were just spent rereading the
- 1:52old chat history in the session. Like
- 1:54that's a huge waste. Now yes, the
- 1:56argument has to be made that well, it
- 1:57needs the context and it needs to
- 1:59understand what we're doing, but still,
- 2:0098.5% is crazy. So take a quick look at
- 2:03this graphic here. Along the x-axis, we
- 2:05have message number, and as it
- 2:07increases, you can see that we have our
- 2:09per message cost and our cumulative
- 2:10tokens increasing, but it's not linear.
- 2:13It's basically each message is rereading
- 2:15all of the past ones, and it has to
- 2:16count that in. So message one could be
- 2:18500, message 30 could be 15,500, which
- 2:21is 31 times more. And then after 30
- 2:23messages, you might already be at almost
- 2:25a quarter million cumulative tokens. Now
- 2:27on top of all of your own messages,
- 2:29Claude will also reload your Claude.md,
- 2:31your MCP servers, your system prompts,
- 2:33your skills, your files on every single
- 2:35turn. And this is invisible overhead,
- 2:37but it is constantly dripping into your
- 2:39context and your tokens. And a really
- 2:41important thing to realize is that
- 2:42bloated context doesn't just cost you
- 2:43more money, but it also produces worse
- 2:45output. So you're paying more and you're
- 2:47getting less. There's this phenomenon
- 2:48called lost in the middle, which
- 2:50basically says that models are paying
- 2:52the most attention in the beginning of
- 2:53your session and kind of at the end. So
- 2:55all that stuff in the middle of your
- 2:56session, kind of in this dip, is getting
- 2:58ignored. All right. So now that we kind
- 3:00of understand a little bit more about
- 3:01how Claude code works and how tokens
- 3:03work, let's move into the hacks. We're
- 3:04going to start here with tier one hacks.
- 3:07These are the ones that are going to be
- 3:08super easy to implement and everyone
- 3:09should be able to understand. So we've
- 3:11got nine of these. Number one is to
- 3:13start fresh conversations. Use {slash}
- 3:16clear between unrelated tasks. Don't
- 3:18carry context about topic A into a
- 3:20conversation about topic B.
- 3:22So every single message in a long chat
- 3:24is exponentially more expensive than the
- 3:25same message in a fresh chat. So this
- 3:28one habit is the number one thing that
- 3:30extends your session life, and it's
- 3:32pretty obvious based on what we just
- 3:33talked about, so that's why this was
- 3:35number one.
- 3:36Okay, number two is to disconnect MCP
- 3:39servers. Every single connected MCP
- 3:41server loads all of its tool definitions
- 3:43into your context on every message. This
- 3:45is another source of completely
- 3:46invisible tokens that might just be
- 3:48eating up and eating away. So one server
- 3:50alone might be something like 18,000
- 3:52tokens per message. So run MCP at the
- 3:55start of each session and disconnect the
- 3:57ones that you don't need. And better
- 3:59yet, if you're able to find CLIs for
- 4:01something, so for example, rather than
- 4:02having the Google Workspace or Google
- 4:04Calendar MCP server, which eats a lot of
- 4:06tokens, just use the Google Workspace
- 4:09CLI. It's faster, it's cheaper, and I
- 4:11think the future is moving towards
- 4:13having our agents use CLIs rather than
- 4:15MCPs. All right, number three, batch
- 4:18prompts into one message. Three separate
- 4:20messages cost three times what one
- 4:22combined message costs because of the
- 4:24way that tokens work, right? Instead of
- 4:26summarize this as one message and then
- 4:28now extract the issues, now suggest a
- 4:30fix, send it all in one prompt. If
- 4:32Claude gets something slightly wrong,
- 4:34edit your original message and
- 4:35regenerate instead of sending a full
- 4:37follow-up correction. Follow-ups stack
- 4:39onto history permanently, while edits
- 4:41replace the bad exchange entirely. Now I
- 4:43will say, there is an argument to be
- 4:45made here that potentially doing it this
- 4:48way where you're doing task one, task
- 4:50two, then task three, might actually be
- 4:52better output quality. I think it
- 4:54depends on the actual use case.
- 4:55Basically, the idea would be if you can
- 4:57give AI one specific task at a time,
- 4:59it's going to be better because it's
- 5:00more specialized and it's more focused.
- 5:02But this is definitely something that
- 5:03you should be aware of. Okay, number
- 5:04four is to use plan mode before any real
- 5:07task. This lets Claude map out the
- 5:09approach, ask you the right questions,
- 5:11and it prevents the single biggest
- 5:12source of token waste, which is just
- 5:14having Claude go down the wrong path,
- 5:15writing code, and then basically
- 5:17everything that it just did, you have to
- 5:19basically like scrap and redo. It's just
- 5:20a huge waste of time and tokens. So you
- 5:22can add something like this to your
- 5:23Claude.md. Do not make any changes until
- 5:25you have 95% confidence in what you need
- 5:27to build. Ask me follow-up questions
- 5:29until you reach that confidence level.
- 5:31This is something that I'm putting into
- 5:33all of my Claude.mds when I am having it
- 5:35help me build things. Number five, we
- 5:37have run {slash} context and {slash}
- 5:39cost. {slash} context shows you exactly
- 5:41what's eating your tokens right now. So
- 5:43your conversation history, your MCP
- 5:44overhead, loaded files, stuff like that.
- 5:47And {slash} cost shows your actual token
- 5:49usage and estimated spend for that
- 5:51current session.
- 5:52Most people have no idea where their
- 5:53tokens are going, and these two commands
- 5:55make the invisible visible. Because if
- 5:58you don't actually know that you're
- 5:59bleeding because of MCPs, then how would
- 6:01you be able to fix that? So when you run
- 6:03{slash} context, this is what it will
- 6:04look like. It'll basically give you a
- 6:06screenshot of how many tokens you're at,
- 6:08what is the cap, and it will estimate
- 6:10based on the different categories. And
- 6:12what I did here is this was ran in a
- 6:14completely fresh session, no chats. So
- 6:16what that tells me is, okay, before I
- 6:18even talk to Claude, I'm already down
- 6:2051,000 tokens because of things like the
- 6:23system prompt, the system tools, my
- 6:24custom agents, my skills, memory files,
- 6:27and here I've actually cleared out all
- 6:29the MCPs, so there wasn't anything in
- 6:30there, but those can, like I said,
- 6:32completely bloat up your tokens right
- 6:34from the get-go. Okay, number six is to
- 6:36set up a status line. This kind of goes
- 6:38hand in hand with having more
- 6:40visibility. You only actually see this
- 6:41in your terminal though, so you will
- 6:42have to do it there. Um and it basically
- 6:45lets you see what's going on. So right
- 6:47here you can see in my terminal, I've
- 6:48got this set up so that I can see the
- 6:49model I'm using. I can see a visual kind
- 6:52of progress bar of my usage, and then I
- 6:54can see
- 6:55uh 5% of my whole 1 million context
- 6:59window, and I can see 52,000 tokens out
- 7:01of 1,000,000, which is a million. And
- 7:03just to clarify, this isn't my session,
- 7:05like my five-hour session. This is
- 7:07basically just indicating that I'm 5% of
- 7:08the way or 52K out of 1,000K. So all you
- 7:12have to do is in Claude code in the
- 7:13terminal, do {slash} status line, and
- 7:15explain that you want to replicate this
- 7:17setup. Number seven is just super
- 7:19simple, but keep your dashboard open.
- 7:20Same thing with visibility. You might
- 7:22run into issues with your limit and just
- 7:25get hit out of nowhere. But if you have
- 7:26it pulled up next to you or you have it
- 7:28ready so that you can switch into that
- 7:29tab and you know, check every 20-40
- 7:31minutes, then you're going to be able to
- 7:33pace yourself a little bit better. You
- 7:34could even set up a automation to
- 7:36basically check in on it every 30
- 7:38minutes and send you like a text or a
- 7:40Slack message to say, "Hey, by the way,
- 7:42you're getting near your usage." All
- 7:45right, so number eight, we have be smart
- 7:46with pasting. Before you drop a document
- 7:49or a file or something large, just ask
- 7:51yourself, "Does Claude need to read this
- 7:53whole thing?" Sometimes it does,
- 7:54sometimes it needs that full context,
- 7:56but sometimes it just needs one section
- 7:58or one page. So if the bug is in one
- 8:00function, then paste just that function.
- 8:02Or if it just needs the context of one
- 8:03little paragraph, just paste that.
- 8:05Claude needs to be precise about what it
- 8:06reads, but you also need to be precise
- 8:08about what you feed it. And number nine,
- 8:10our last tier one hack, is to actually
- 8:12watch Claude code work. Don't just fire
- 8:14off a prompt and walk away or switch
- 8:16tabs. Watch what Claude is doing,
- 8:18especially on longer tasks. And this is
- 8:20because if you actually sit and watch
- 8:21it, sometimes you'll realize it's going
- 8:23down the wrong path. Sometimes it gets
- 8:24stuck in its own loops, rereads the same
- 8:26files, things like that. So if it's
- 8:29doing that, you might as well just stop
- 8:31it right there. Kind of the same idea as
- 8:32plan mode. Why would you let it go down
- 8:34the wrong path, waste all your tokens,
- 8:35and then just have to scrap it all? In a
- 8:37bad loop, 80% of the tokens are being
- 8:39used producing zero value. So if you're
- 8:42able to just watch your session run
- 8:44until you know it's going down the right
- 8:46path, it could save you thousands of
- 8:47tokens. All right, let's kick it up a
- 8:49little bit. Let's move into our tier two
- 8:50hacks, and for these ones we have five
- 8:53of them. So number one is to keep your
- 8:55Claude.md file lean. Place it in your
- 8:57project root, whether that is globally
- 8:59or in in local projects. Claude auto
- 9:02reads it at the start of every single
- 9:03chat as system context. So keep it under
- 9:06200 lines. Include things like your tech
- 9:08stack, your coding conventions, your
- 9:09build commands, the 95% confidence rule,
- 9:12only the most important things, and you
- 9:14need to treat this like an index. Route
- 9:16to where more data lives. And it's a
- 9:18complete mindset shift. This file
- 9:20basically just tells Claude code, "Where
- 9:21is everything that it needs and what to
- 9:23do every single time?" So it can point
- 9:25to files that are huge, but that way it
- 9:26just is, "Okay, I don't need this right
- 9:28now, but if I do need this, I know
- 9:29exactly where to look." And because it
- 9:31knows exactly where to look, it's not
- 9:32going to waste time or tokens searching
- 9:34through and reading other files. It's
- 9:35just able to grab it right there by the
- 9:37file name. And the reason I say this is
- 9:38a mindset shift because you should be
- 9:40doing this with other things, not just
- 9:41your Claude.md, with your skills or with
- 9:44your you know, master reference guide
- 9:45sheets. I saw someone talking about how
- 9:47they created an index that's super super
- 9:49lean, and it shows Claude code exactly
- 9:51where to go in the Claude code
- 9:53documentation. So if it needs help with
- 9:55something related to Claude code, it
- 9:56doesn't have to search through the whole
- 9:58database, it can just say, "Okay, here's
- 9:59my index file. I know exactly which URL
- 10:01to look up at." Super simple. You want
- 10:03to keep this lean and trim it all the
- 10:04time. It's always a work in progress
- 10:06because every single chat, not just like
- 10:08your session, every single message
- 10:10Claude.md gets read. So, if your
- 10:11Claude.md file is 1,000 lines, every
- 10:14single time you shoot off a message,
- 10:15even if you just say hi,
- 10:17the whole thing's going to get read.
- 10:18Okay, number two here is to be surgical
- 10:20with file references. Don't just say
- 10:22something like, "Here's my whole repo,
- 10:24go find the bug." Say something more
- 10:26like, "Check the verify user function
- 10:27inside the off.js file." Or you can also
- 10:31use @filename to point at specific files
- 10:34instead of once again letting Claude
- 10:35explore freely. The whole idea of being
- 10:38specific and routing. All right, so
- 10:39number three, I'm saying to compact at
- 10:41around 60% capacity. Auto compact
- 10:44triggers at like 95% by which point your
- 10:46context is already pretty degraded. So,
- 10:48run/context to check your capacity
- 10:50percentage or you should have the status
- 10:51line set up and at about 60% just run
- 10:54the /compact with specific instructions
- 10:56on what it should actually be
- 10:57preserving. After three to four compacts
- 10:59in a row, the quality does start to
- 11:01degrade. So, at that point, once you've
- 11:03done three or four, just get a session
- 11:05summary, /clear, give the session
- 11:08summary back, and then keep going. All
- 11:10right, so number four, short breaks are
- 11:11actually costing you. Claude code uses
- 11:14prompt training to avoid reprocessing
- 11:16unchanged context, but the cache has a
- 11:185-minute timeout. So, if you step away
- 11:20and you come back and it's been longer
- 11:21than 5 minutes, your next message
- 11:23reprocesses everything from scratch at
- 11:25full cost. And that is why some people
- 11:27feel like their usage just randomly
- 11:29spikes if they might have, you know,
- 11:30stepped away and came back. So, if
- 11:32you're going to do that, just consider
- 11:33doing a /compact or a /clear before you
- 11:36step away. All right, number five,
- 11:37command output bloat. When Claude runs
- 11:40shell commands, the full output enters
- 11:42your context window. So, if you have a
- 11:45command then it comes back with 200
- 11:46commits or, you know, just tons and tons
- 11:48of data, then all of that is tokens that
- 11:51get sent to your model. So, really the
- 11:54takeaway here is to be intentional about
- 11:56what you let Claude run. If you know in
- 11:58a certain project that it doesn't need
- 11:59to use certain commands, then you can go
- 12:01ahead and in that project deny those
- 12:03permissions. And this is another one
- 12:05that seems like it's invisible because
- 12:07when it runs like a bash or, you know,
- 12:08certain commands, it basically just has
- 12:11like one line and you don't actually see
- 12:12all of the tokens that it has, you know,
- 12:14sent there. All right, so sitting here
- 12:17editing this video and there's just one
- 12:18more thing that I wanted to get off my
- 12:20chest and it's basically about hitting
- 12:22your limit. And, you know, the goal of
- 12:24this video and your goal should be to
- 12:26optimize so that you don't hit your
- 12:27limit, but I don't think that you should
- 12:29associate hitting your limit with like
- 12:31it shouldn't be a negative connotation
- 12:33because ultimately if you're doing a lot
- 12:35of these hacks and you are not just like
- 12:38being wasteful with tokens, then hitting
- 12:40your limit is actually a good thing if
- 12:42you think about it because it means that
- 12:44you are using this tool so much and I
- 12:45think that's what you want to be. I
- 12:47think you want to be a power user of
- 12:49this tool to the point where it's like
- 12:51got to wait again. And, you know,
- 12:53waiting sucks, but
- 12:55people that are using it so much are
- 12:57going to be so much more productive and
- 12:59so much farther ahead than people who
- 13:00are never hitting their limits, you
- 13:02know, not make not getting their their
- 13:04money's worth and not truly
- 13:08getting the leverage that you are now
- 13:09getting. So, anyways, quick little raw
- 13:11rant there, but I think it's an
- 13:13important mindset shift to have, you
- 13:14know?
- 13:15Just something to think about. All
- 13:17right, so we're moving on to tier three
- 13:18now. I hope you guys feel like you
- 13:20already have a lot of things that you
- 13:21want to implement and these ones are
- 13:23getting a little crazier as well. So,
- 13:24we've got four of these here and I've
- 13:25got a few bonus ones also, but number
- 13:27one is to pick the right model. So,
- 13:29Sonnet for your default most coding
- 13:31work, Haiku for sub-agents, formatting,
- 13:33simple tasks, Opus for deep
- 13:35architectural planning and only when
- 13:36Sonnet wasn't enough. Try to keep this
- 13:38under 20% of usage or unless you just
- 13:41really, really need it for that project.
- 13:42Now, a little tip here is when you have
- 13:44a huge code base and you want to do
- 13:45certain things like maybe a review, then
- 13:47try bringing in Codex. There is an
- 13:49official plugin now and I made a video
- 13:51about this. I'll tag that right up here,
- 13:52but you could basically have, you know,
- 13:53Opus and Sonnet working together to
- 13:55build you, you know, a project or a code
- 13:57base and then you could just bring in
- 13:58Codex to actually review everything. And
- 14:00that way you're saving yourself on the
- 14:02Claude tokens. The next one, number two
- 14:05here, is the cost of sub-agents. Agent
- 14:07workflows use roughly seven to 10 times
- 14:09more tokens than a standard single agent
- 14:11session. Now, why is that? Because they
- 14:13wake up with their own full context and
- 14:15it's a separate instance. So, they
- 14:17basically have to reload everything when
- 14:19you start up the new session. All of
- 14:21those files, all of the system tools,
- 14:23like everything like that. Now, what you
- 14:24can do though, which is helpful, is to
- 14:26delegate to sub-agents for one-off
- 14:27tasks, especially if you want that
- 14:29one-off task to use Haiku. So, maybe you
- 14:31need to process a lot of information or
- 14:33maybe you need to do a ton of research
- 14:34and get just like a summary back. Now,
- 14:36yes, tokens are still tokens no matter
- 14:38what at the end of the day, but if you
- 14:40can make 80% of your tokens a cheaper
- 14:42model rather than 80% of your tokens an
- 14:44expensive model, then you're going to be
- 14:46saving money.
- 14:47And then, of course, agent teams are
- 14:48cool. Um sometimes I really do actually
- 14:51like them and it helps me get more
- 14:53higher quality outputs, but they're
- 14:55very, very expensive. So,
- 14:57try to use them very sparingly. All
- 14:58right, so number three is to understand
- 15:00peak hours. So, we just talked about at
- 15:02the beginning how they've adjusted how
- 15:04fast your 5-hour session window drains
- 15:06based on demand during the peak hours,
- 15:08which are 8:00 a.m. to 2:00 p.m. Eastern
- 15:11Time on weekdays.
- 15:13But off-peak, this is when your usage is
- 15:15kind of either normal or it lasts a
- 15:17little longer and these are afternoons,
- 15:19evenings, weekends. So, if you actually
- 15:21think about this strategically, maybe
- 15:23you want to make sure that you're
- 15:23running big refactors or multi-agent
- 15:26sessions or big projects during off-peak
- 15:28hours only. Otherwise, you're going to,
- 15:30you know, drain right through that peak
- 15:32session. And on top of this, we'll call
- 15:34this a little hack 3.5, which is the one
- 15:36I kind of alluded to earlier when I
- 15:38said, "Hey, just keep open your Claude
- 15:40account so you can see your usage at all
- 15:41times." If you're near a reset and you
- 15:43have room left in your allocation, then
- 15:44go heavy. Try to hit that usage limit
- 15:46before it resets. Get your money's
- 15:48worth. Let your agents go loose at that
- 15:50point. And on the other side, if you're
- 15:52getting near your limit, but you still
- 15:53have lots of time, then step away. This
- 15:56is your time to take a break, take a
- 15:57walk, make some lunch, come back with a
- 16:00full budget instead of burning the last
- 16:025% on something small and getting stuck
- 16:04mid-task and having to just kind of, you
- 16:05know, lose that flow state that you
- 16:07might have been in. Okay, number four,
- 16:10your system's constitution,
- 16:12which is Claude.md. This should contain
- 16:14stable decisions, architecture rules,
- 16:16and progress summaries. Think of it like
- 16:18the source of truth that makes every
- 16:19prompt shorter and shorter. Save
- 16:21decisions, not conversations. Every
- 16:23architectural call that you store there
- 16:25is a paragraph that you never have to
- 16:26type again. So, this builds on top of
- 16:28the way that you were thinking about it
- 16:29back in tier one. You can add rules in
- 16:32there that basically tell it, "Hey, I
- 16:33want you to help me make sure I'm being
- 16:36smart about tokens. Use sub-agents for
- 16:37any exploration or research. If a task
- 16:39needs three plus files or multi-file
- 16:41analysis, spawn a sub-agent and only
- 16:43return summarized insights. Spawn that
- 16:45sub-agent in Haiku." And here's a little
- 16:47prompt that I have at the bottom of my
- 16:49Claude.md and I will say before I read
- 16:51this out, you have to be careful because
- 16:53when you make a file like this, um kind
- 16:55of self-learning or self-evolving, you
- 16:57have to check on it frequently because
- 16:59you don't want it to accidentally get
- 17:00too bloated. But here I said, "Applied
- 17:02learning. When something fails
- 17:03repeatedly, when Nate has to re-explain
- 17:05or when a workaround is found for a
- 17:07platform tool or limitation, add a
- 17:09one-line bullet here. Keep each bullet
- 17:11under 15 words, no explanations. Only
- 17:13add things that will save time in future
- 17:14sessions." And then it's got some
- 17:15bullets. Now, I'm not saying this is the
- 17:17most optimal prompt, but I think this
- 17:19sort of system of having your Claude.md
- 17:22actually learn and continuously think
- 17:24about how it can save you time and
- 17:26tokens is a good idea to play with. All
- 17:28right, so I know that we just went
- 17:29through a ton of stuff. This whole slide
- 17:31deck will be available for download for
- 17:33free in my free school community. The
- 17:35link for that will be down in the
- 17:36description. But right now, what you
- 17:37should go do are these things. Go
- 17:39run/context. See what it looks like. Go
- 17:42to some of your active sessions,
- 17:43run/cost. Status line, make sure it's
- 17:45showing your model, your context
- 17:47percentage, and your token count. Make
- 17:48sure you pull up your Claude usage
- 17:50dashboard so you can see your remaining
- 17:52allocation and what time it resets.
- 17:54Disconnect unused MCP servers. Start
- 17:57complex tasks in plan mode. Use /clear
- 17:59when you're switching to an unrelated
- 18:00task. Manually compact at 60% context.
- 18:04Batch your multi-step instructions into
- 18:06single messages and schedule heavy
- 18:08sessions for off-peak hours and really
- 18:10just be mindful about the actual timing.
- 18:12So, I wanted to kind of leave you guys
- 18:13with one, maybe two messages. The first
- 18:15thing is just the idea that there is a
- 18:18balance between quality and cost. And
- 18:21so, that's kind of a game that you have
- 18:22to play a little bit and sometimes you
- 18:23do have to go for the higher quality,
- 18:24which ultimately is going to cost you
- 18:26more money and that's just the way it
- 18:27works. But the other thing is just to
- 18:28keep it simple and think about what we
- 18:31talked about at the beginning of this
- 18:32video, how tokens actually work, how
- 18:34Claude code actually charges you. Most
- 18:36people don't need a bigger plan, they
- 18:37need to stop resending their entire
- 18:39conversation history 30 times when you
- 18:41could just send it, you know, five
- 18:42times. It's not a limits problem, it's a
- 18:44context hygiene problem. But anyways,
- 18:47that is going to do it for this one. If
- 18:48you guys enjoyed or you learned
- 18:49something new, please give it a like, it
- 18:50helps me out a ton. And as always, I
- 18:52appreciate you guys making it to the end
- 18:53of the video. I'll see you all in the
- 18:54next one.
- 18:55Thanks, everyone.
About this transcript
This page contains the full transcript of 18 Claude Code Token Hacks in 18 Minutes by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 4,251 words across 609 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.