Never Hit Claude Usage Limits Again. 15 Best Hacks (2026) – Transcript

00:00I bet that if you showed me how you're using Claude right now, I could find at least four or five different ways that you could get more out of Claude using fewer tokens. This is a question I get asked all the time. How can I avoid hitting my Claude limits?

00:14So, in this video, I'm going to show you 15 ways that you can push Claude further while spending fewer tokens. And I'm going to break these into three categories. The first category is being smart about which model, which effort level, and which level of reasoning that you're using in Claude.

00:29And this isn't as simple as just use Sonet for everything. I'm going to show you techniques that work for Opus or Fable or any model. The second category is one and done changes you can make to your code setup so that it is always more efficient. And these are tweaks and optimizations you can make once that will work forever.

00:47And the final category is habits that I think you should get into when you're working with AI to get better results with fewer tokens. So, let's get started. Pretty much everything I'm going to show you here applies to Cloud Chat, Cloud Co-work, and Cloud Code. So, there is something in here for everyone.

01:04So, let's start with model choice, the most obvious way to save tokens. Now, if you're on a $200 a month plan, you can probably get away with using Opus most of the time, and I doubt you're going to hit your limits if you apply the rest of these techniques. But if you're on a $20 a month or a $90 a month plan, chances are you have been hitting limits. So I recommend to people that they use Sonnet as their default model and then they use Opus when they're doing more strategic or more planning types of work. People

01:31have asked me, well, how do I know when I should use Opus? And the question I suggest asking yourself is, am I working on something that I will reuse in the future? So, if I'm creating a skill, if I am building some sort of tool or a very important piece of strategic work that I'm going to refer to over and over again, I would bump up to Opus. But if I'm just doing a oneanddone piece of work, if I'm just doing routine, you know, summarization or getting, you

01:58know, my latest transcripts from Granola, then that is a Sonnet task. We do not need super intelligent models to do a lot of our basic AI work now. and Sonnet is really a fantastic model at a fraction of the price of Opus and Fable.

02:13If you're planning out a big piece of work that you want Claude to complete, then I would also suggest using Opus to write the plan and then handing that plan to a new chat with Sonnet. I'll show the current pricing on screen right now, but this changes all the time. And as a general rule, I do think that using Sonnet as your default is the best way to push your subscription as far as possible. So, let's talk about effort level. This is a setting that Anthropic define as controlling how many tokens

02:39Claude uses when responding. And by default, you'll see the effort level set to high. There are actually five effort levels: low, medium, high, extra high, and max. Now, I suggest that you almost never go near extra high or max because you probably don't need them unless you're working on something absolutely mission critical. But if you find that you're often hitting uh usage limits,

03:06you can pull the effort level down from high to medium and this means that COD will just think a little bit less before it gets back to you. Again, if you're doing routine work, if you're just summarizing something, then you can probably get away with setting the effort level to medium and your response will be almost identical to if it was high. One important thing is that if you change the effort level during a conversation, then Claude will have to reread the entire conversation. So try not to change it in the middle of a

03:34chat. If you're starting to run out of tokens, I suggest starting a new conversation and taking that effort level down from high to medium. Another benefit of this is that if you take the effort level down from high to medium, you will actually get faster responses from Claude. So if you're working to a deadline or you feel like you're running out of time, then medium effort is probably going to save you time. The third thing we can do is switch off thinking mode or sometimes known as extended thinking mode. And in

04:01Anthropic's documentation, they actually list this as a very effective way to reduce token usage. So thinking mode is basically Claude reflecting on its own answers and reasoning a little bit more about everything before it comes back to you with an answer. This is helpful if you are doing very important strategic work. But again, if you're doing routine work, if you're just summarizing or synthesizing information with an LLM, you probably don't need thinking mode

04:27on. And I would highly recommend that you turn that off. And you can find this by going into the model selector, then coming down to thinking. So if you're making a pretty simple request of Claude and it's something that maybe needs one or two steps before it can be completed, then you might not need thinking switched on. But if you're asking Claude to go out and use multiple connectors and then look at multiple files and then think about the trade-offs of a decision, that is when you need to keep thinking switched on. All right. So now

04:56I'm going to show you seven things you can do once that will immediately reduce the amount of tokens that you find yourself using. The first one is keeping your cloud.md or project instructions pretty slim. If you don't know cloud.md is a file that gets loaded into the start of every conversation in co-work or code and in chat if you're using projects you can define custom instructions. In any case this is a big set of instructions that get loaded into

05:23the start of every conversation. If you've been working with Claude for a long time, this set of instructions might have become quite bloated and so every chat you have will be bringing in quite a long set of instructions. It is worth just looking at your cloud. MD and asking Claude, "How could we tidy this up without having any reduction in effectiveness?" So, if you want to tidy this up, here's a simple prompt that you can use in co-work.

05:48Simply ask Claude, "Read my Claude.MD MD and look for ways that we could make this more concise without losing any of its effectiveness. And if you're in Claude chat, then just copy your project instructions, paste them into a new session, and ask Claude the same question. Okay, the next thing we can do is ask Claude to be more concise with its answers. And the reason for this is that output tokens actually cost five

06:14times more than input tokens. So, we really want to keep Claude's outputs as compact and succinct as possible. With some of the more recent models, this has become quite problematic. Opus 5 and even Fable have become quite verbose.

06:29So, we want to tell Claude to be concise in its responses with us. There's a few ways we can achieve this. The simplest way is to go into settings in Claude and then in your general preferences page, add some instructions to the global instructions file. This affects Claude chat and Claude cowwork and ask Claude to be concise in its answers. You can ask it to use bullets. You can ask it to avoid padding out its answers. And you can ask it to cut the fluff. I've got a

06:57link in the description that basically shows you the instructions I use that keep Claude pretty concise in its responses. There's a really fun way of doing this and it's called Caveman mode. This started out as a plugin for Claude Code that the creator says reduced tokens by 65%. And it does this by asking Claude to respond like a caveman.

07:16So you get very short responses from Claude. And you can install this as a plugin in Claude Code or there's a set of instructions again that you can paste into your uh preferences or paste into your Clauded file that will make Claude respond like a caveman. Another way you can do this is a skill called I have ADHD. And this is a skill that just forces Claude to respond in very clear,

07:41very structured, very short answers. Now, this is a skill, but again, you can take the concepts of this skill and drop it into your preferences or your Claude.md. So again, all of this is linked in the description, but if you do any of these, then you will notice that Claude is using fewer tokens because it's giving you more concise answers.

08:04often answers that have a higher signal to noise ratio. The third thing we can do is to give Claude a lot of context so that you don't have to keep pasting the same information into your conversations. If I started an entirely fresh Claude account, which knew nothing about me, all of my conversations would have to include so much information that Claude would have to process just to give me a helpful reply. However, if we

08:30have a second brain folder, for example, or just lots of context files that Claude can read from, then we don't need to paste stuff in nearly as often, and we will see that Claude is using fewer tokens. The best way to store all of your context is in something called a markdown file. This is a tiny file size that just lets Claude read structured text. So if you have PDFs, big documents, you want to be converting

08:57those to markdown files because they use something like 90% fewer tokens. And this is a really easy way to reduce the amount of tokens that you're using in Claude. The next technique we can apply is a slightly more advanced one, and this is for if you're using Claude cowwork or Claude code in quite a big folder with lots of files. Now one way that Claude burns through tokens is just searching through files to find the right information. And if we want to

09:24reduce this token usage, then we can structure our folder in a very agentfriendly way. There is a researcher called Andre Carpathy who a few months ago released something brilliant called the LLM wiki. Now, I'm not going to get into all the specifics of how that works right now, but it's basically a framework for keeping your second brain or your main sort of context folder organized in a very logical way that makes it very easy for agents to find

09:50information quickly. For example, we will ask our agent to keep an index file of everything in the folder so that it can easily look things up. We will put summaries at the top of every file so that an agent can read the summary and then decide if that's the right file or the wrong file. And we will link all of these files to other relevant files, which just makes it easier for Claude to find information. Now, when I was applying this to my own second brain a few months ago, I ran some tests and

10:17there was one prompt where I asked Claude to find a very specific piece of information that took 10 different searches until it could find exactly what it was looking for. And I started, you know, a fresh conversation and just said, "Go and find this." When I applied the LLM wiki principles, it was actually able to find that in a single search. It went through 3D index. It found the right file straight away. And so it used far fewer tokens just trying to return information to me. So if you want to

10:45apply this to your folder, there's a link in the description. Just go and copy that prompt, feed it into Claude, and you will almost immediately notice that Claude finds things faster and uses fewer tokens to find and update different files in your folder. One really effective thing we can do is create skills for Claude. Now, if you haven't come across skills, they are a set of instructions that tell Claude how to carry out a particular task. Without

11:11skills, every time we want Claude to do something, it has to figure out how to do that task. For example, if I had a skill that was my morning briefing and it pulled data from my CRM and my granola and Stripe without a skill, every time I asked Claud to give me a morning briefing, it would have to go and figure out where the information was and how to get it and the briefings might be in a completely different format every day. That is quite an

11:38expensive way of working with Claude. Instead, if we create a skill, then Claude just immediately follows the instructions and we're using far fewer tokens. Now, there's a really powerful feature in co-work called skill creator, and it's slightly hidden, but what it does is it interviews you about a piece of work or it takes a SOP or a process document that you've already got. And it has a first attempt at creating a skill.