**Never Hit Claude Usage Limits Again. 15 Best Hacks (2026) — Transcript & Summary | SozAI**
Source: https://sozai.app/transcript/never-hit-claude-usage-limits-hacks/

Learn 15 hacks to maximize Claude AI usage and avoid hitting token limits by optimizing model choice, effort level, and workflow habits.

## Key Takeaways

- Choosing the appropriate Claude model and settings can significantly extend your usage limits.
- Reducing effort level and disabling thinking mode are effective ways to save tokens without sacrificing output quality for routine tasks.
- Concise instructions and responses help minimize token consumption, especially since output tokens cost more than input tokens.
- Organizing workflows with reusable skills, external context, and sub-agents optimizes token usage and efficiency.
- Regularly reviewing and trimming your Claude.md or project instructions prevents bloating and excessive token use.

## What the video covers

- Use the right Claude model for the task: Sonnet for routine work and Opus for strategic or reusable tasks to save tokens and cost.
- Adjust effort levels from high to medium to reduce token use and speed up responses without losing much quality.
- Turn off thinking mode for routine tasks to avoid unnecessary token consumption.
- Keep Claude.md or project instructions concise to reduce token load at conversation start.
- Instruct Claude to be concise in responses, using bullets and cutting fluff to minimize output tokens which are costlier.
- Use 'caveman mode' to drastically shorten responses and save tokens, especially in Claude Code.
- Leverage external context files or folders to avoid repeatedly pasting information into Claude.
- Create and use reusable skills to streamline workflows and reduce token use over time.
- Keep connectors off by default and enable only when necessary to avoid extra token costs.
- Use sub-agents (parallel conversations) to handle tasks efficiently and reduce token usage in main threads.

## Chapters

1. 00:00 Introduction and Overview
2. 00:56 Applicability Across Claude Versions
3. 01:40 Choosing the Right Claude Model
4. 02:34 Understanding and Adjusting Effort Levels
5. 03:32 Managing Effort Level Changes During Conversations
6. 04:28 Turning Off Thinking Mode to Save Tokens
7. 05:31 Optimizing Claude.md and Project Instructions
8. 06:17 Encouraging Concise Responses from Claude
9. 07:13 Using Caveman Mode for Token Reduction
10. 08:29 Leveraging External Context and Skills for Efficiency

Answers

## Questions about this video

How can I avoid hitting Claude's token usage limits?

You can avoid hitting Claude's token limits by choosing the right model for your task, lowering the effort level, turning off thinking mode for routine tasks, and keeping your project instructions concise.

When should I use the Opus model instead of Sonnet?

Use Opus for strategic, reusable, or important work that you will refer to multiple times, and use Sonnet for routine, one-off tasks like summarization to save tokens and cost.

What is the impact of changing effort levels during a conversation?

Changing the effort level mid-chat causes Claude to reread the entire conversation, increasing token usage, so it’s better to start a new conversation when adjusting effort levels.

## Full Transcript — Download SRT & Markdown

00:00

Speaker A

I 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.

00:09

Speaker A

This is a question I get asked all the time: How can I avoid hitting my Claude limits?

00:14

Speaker A

So in this video, I'm going to show you 15 ways that you can push Claude further while spending fewer tokens.

00:20

Speaker A

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.

00:30

Speaker A

And this isn't as simple as just use Sonnet for everything. I'm going to show you techniques that work for Opus or Fable or any model.

00:36

Speaker A

The second category is one-and-done changes you can make to your Claude setup so that it is always more efficient.

00:43

Speaker A

And these are tweaks and optimizations you can make once that will work forever. And 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.

00:56

Speaker A

So let's get started. Pretty much everything I'm going to show you here applies to Claude Chat, Claude Cowork, and Claude Code.

01:02

Speaker A

So there is something in here for everyone. So let's start with model choice, the most obvious way to save tokens.

01:08

Speaker A

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.

01:16

Speaker A

But if you're on a $20 a month or a $90 a month plan, chances are you have been hitting limits.

01:22

Speaker A

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 type of work.

01:30

Speaker A

People have asked me, well, how do I know when I should use Opus?

01:34

Speaker A

And the question I suggest asking yourself is, am I working on something that I will reuse in the future?

01:40

Speaker A

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.

01:51

Speaker A

But if I'm just doing a one-and-done piece of work, if I'm just doing routine summarization or getting my latest transcripts from Granola, then that is a Sonnet task.

02:02

Speaker A

We do not need super intelligent models to do a lot of our basic AI work now.

02:08

Speaker A

And Sonnet is really a fantastic model, at a fraction of the price of Opus and Fable.

02:13

Speaker A

If 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.

02:22

Speaker A

I'll show the current pricing on screen right now, but this changes all the time.

02:26

Speaker A

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.

02:34

Speaker A

So let's talk about effort level. This is a setting that Anthropic defines as controlling how many tokens Claude uses when responding.

02:42

Speaker A

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.

02:53

Speaker A

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.

03:00

Speaker A

But if you find that you're often hitting usage limits, you can pull the effort level down from high to medium.

03:09

Speaker A

And this means that Claude will just think a little bit less before it gets back to you.

03:14

Speaker A

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.

03:25

Speaker A

One important thing is that if you change the effort level during a conversation, then Claude will have to reread the entire conversation.

03:32

Speaker A

So try not to change it in the middle of a chat. 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.

03:42

Speaker A

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.

03:49

Speaker A

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.

03:55

Speaker A

The third thing we can do is switch off thinking mode, or sometimes known as extended thinking mode.

04:01

Speaker A

And in Anthropic's documentation, they actually list this as a very effective way to reduce token usage.

04:08

Speaker A

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.

04:16

Speaker A

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 on.

04:28

Speaker A

And I would highly recommend that you turn that off. And you can find this by going into the model selector and coming down to thinking.

04:36

Speaker A

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.

04:44

Speaker A

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.

04:56

Speaker A

All right, so now I'm going to show you seven things you can do once that will immediately reduce the amount of tokens that you find yourself using.

05:04

Speaker A

The first one is keeping your Claude.md or project instructions pretty slim. If you don't know, Claude.md is a file that gets loaded into the start of every conversation in Cowork or Code.

05:16

Speaker A

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 the start of every conversation.

05:25

Speaker A

If you've been working with Claude for a long time, this set of instructions might have become quite bloated.

05:31

Speaker A

And so every chat you have will be bringing in quite a long set of instructions.

05:36

Speaker A

It is worth just looking at your Claude.md and asking Claude, how could we tidy this up without having any reduction in effectiveness?

05:44

Speaker A

So if you want to tidy this up, here's a simple prompt that you can use in Cowork or Code.

05:49

Speaker A

Simply ask Claude, read my Claude.md and look for ways that we could make this more concise without losing any of its effectiveness.

05:58

Speaker A

And if you're in Claude Chat, then just copy your project direct instructions, paste them into a new session, and ask Claude the same question.

06:05

Speaker A

Okay, the next thing we can do is ask Claude to be more concise with its answers.

06:10

Speaker A

And the reason for this is that output tokens actually cost five times more than input tokens.

06:17

Speaker A

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.

06:26

Speaker A

Opus 5 and even Fable have become quite verbose. So we want to tell Claude to be concise in its responses with us.

06:33

Speaker A

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.

06:44

Speaker A

This affects Claude Chat and Claude Cowork and ask Claude to be concise in its answers.

06:51

Speaker A

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.

06:57

Speaker A

I've got a link in the description that basically shows you the instructions I use that keep Claude pretty concise in its responses.

07:04

Speaker A

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%.

07:13

Speaker A

And it does this by asking Claude to respond like a caveman. So you get very short responses from Claude.

07:20

Speaker A

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 preferences or paste into your Claude.md file that will make Claude respond like a caveman.

07:33

Speaker A

Anoth

07:45

Speaker A

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.

07:53

Speaker A

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, often answers that have a higher signal to noise ratio.

08:09

Speaker A

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.

08:17

Speaker A

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.

08:29

Speaker A

However, if we have 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.

08:40

Speaker A

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.

08:48

Speaker A

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 those to markdown files because they use something like 90% fewer tokens.

09:03

Speaker A

And this is a really easy way to reduce the amount of tokens that you're using in Claude.

09:08

Speaker A

The next technique we can apply is a slightly more advanced one. And this is for if you're using Claude Cowork or Claude Code in quite a big folder with lots of files.

09:17

Speaker A

Now, one way that Claude burns through tokens is just searching through files to find the right information.

09:24

Speaker A

And if we want to reduce this token usage, then we can structure our folder in a very agent-friendly way.

09:31

Speaker A

There is a researcher called Andre Karpathy, who a few months ago released something brilliant called the LLM Wiki.

09:37

Speaker A

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 information quickly.

09:52

Speaker A

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.

09:58

Speaker A

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.

10:05

Speaker A

And we will link all of these files to other relevant files, which just makes it easier for Claude to find information.

10:12

Speaker A

Now, when I was applying this to my own second brain a few months ago, I ran some tests and there 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.

10:28

Speaker A

And I started, you know, a fresh conversation. just said, go and find this. When I applied the LLM wiki principles, it was actually able to find that in a single search.

10:37

Speaker A

It went through the index, it found the right file straight away. And so it used far fewer tokens just trying to return information to me.

10:44

Speaker A

So if you want to apply this to your folder, there's a link in the description.

10:48

Speaker A

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.

11:01

Speaker A

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.

11:11

Speaker A

Without skills, every time we want Claude to do something, it has to figure out how to do that task.

11:18

Speaker A

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 Claude to give me a morning briefing, it would have to go and figure out where the information was and how to get it.

11:35

Speaker A

And the briefings might be in completely different formats every day. That is quite an expensive way of working with Claude.

11:41

Speaker A

Instead, if we create a skill, then Claude just immediately follows the instructions and we're using far fewer tokens.

11:49

Speaker A

Now, there's a really powerful feature in Cowork called Skill Creator, and it's slightly hidden.

11:53

Speaker A

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.

12:00

Speaker A

it has a first attempt at creating a skill. And when it's made that skill, it will ask you for feedback.

12:06

Speaker A

So for example, if I was creating a morning briefing skill, then it would have a first go and send me three briefings for my review.

12:14

Speaker A

And it opens a window asking for your feedback. And I would say, you know, this one is too verbose.

12:20

Speaker A

This one completely missed a section. And this one is about right, actually. And so then Claude takes all that feedback and iterates again.

12:29

Speaker A

And then you give more feedback. And eventually, every single briefing it creates is bang on.

12:34

Speaker A

And that's what you want. This process, as well as trying to give you the best output, also makes Claude look for the most token efficient way to create skill.

12:45

Speaker A

So by using Skill Creator, not only are you packaging up a repeatable piece of work that will save you tokens in the future, but Claude is actually looking for the most efficient way to create that skill for you.

12:57

Speaker A

If you've already got skills, then you can actually say to Claude, look for ways to make this skill more token efficient, and it will set up a test harness and basically run a few tests, measure the token usage, and then tell you which is the best way going forwards.

13:12

Speaker A

So creating skills is great, but it can actually be quite a token intensive process.

13:18

Speaker A

So we might use quite a lot of tokens to create this skill, and then we'll reap the benefits and save all the tokens going forwards every time we use the skill.

13:26

Speaker A

So there's loads of directories out there for finding skills, and you might see people share.

13:30

Speaker A

in them on LinkedIn or Twitter. But I think the best one is skills.sh, which is directory created by the team at Vercel.

13:39

Speaker A

So if you go to skills.sh, you'll see that the most popular skill is actually a skill called find skills.

13:46

Speaker A

And if you install this in Claude, then you can basically ask Claude to go and check whether any skill exists that does what you want it to do.

13:55

Speaker A

So before you spend a lot of tokens creating your own skill, you could go and check whether someone's done this already.

14:01

Speaker A

For example, if you work in sales and you want to create like a pre-call briefing skill that pulls data from your CRM and maybe does a web search and maybe does some enrichment.

14:11

Speaker A

Yes, you could create your own skill. And I do recommend this if you want it to be highly customized.

14:17

Speaker A

Or you could invoke the find skills skill from Vercel. Ask Claude, do any skills exist for getting me ready before a call?

14:26

Speaker A

And you will see that there are lots of sales skills that people have shared already.

14:30

Speaker A

You can look at how many installs they've had to sort of gauge which ones are most popular.

14:34

Speaker A

And you can just drop that into Claude and start using it immediately. You will probably need to tweak it a bit and customize it to work for you.

14:40

Speaker A

But that can save you a lot more tokens than creating a skill from scratch.

14:45

Speaker A

Okay, so now let's talk about some good habits that will definitely save you tokens.

14:49

Speaker A

The first one is quite an easy one. When you go into Claude and look at your connectors, you'll see that you can turn them on and off.

14:56

Speaker A

And by turning off the connectors you're not using, you will act. actually save tokens.

15:02

Speaker A

So this is a really easy one, but keep your connectors off by default and only switch them on when you actually need them.

15:10

Speaker A

And this will use fewer tokens. The next thing we can do is to regularly start fresh conversations.

15:16

Speaker A

A lot of people have one mega conversation going with Claude that keeps getting compacted.

15:23

Speaker A

And they believe that if they keep having this conversation in a single thread, Claude will remember everything.

15:28

Speaker A

And this just isn't true. We really want to store everything important in our context files so that we can just spin up a new chat whenever we want and Claude will be able to find what it needs without needing all that past conversation.

15:42

Speaker A

This saves you tokens because as your context window or as the length of the conversation gets bigger and bigger, Claude needs to spend more tokens to process what you're saying.

15:54

Speaker A

So if you've been having a conversation with hundreds of messages in it and then you ask a question, that will take a lot more tokens than if you were to start a new conversation and ask the question.

16:07

Speaker A

Keep trying to start fresh conversations whenever you can and try to avoid these really big conversations.

16:14

Speaker A

And if you do find that you're having really big conversations quite often, then get into the habit of saving important facts and figures in context instead of relying on the context window.

16:25

Speaker A

Okay, so this habit will not only save you lots of tokens, but it will also save you a lot of...

16:30

Speaker A

headaches and it will avoid this feeling of trial and error that you sometimes have when you're trying to get Claude to do something complex.

16:39

Speaker A

I'm going to take the example of asking Claude to build a landing page. A lot of people, when they approach that problem, would simply give a prompt to Claude that says something like, build me a landing page for this new feature on my website.

16:55

Speaker A

It is called, I don't know, whatever the feature is called. And they would hit enter and sort of wait for Claude to start asking questions.

17:01

Speaker A

And then they would sort of give more information as Claude asked. And Claude would attempt to build this page and it would look terrible.

17:08

Speaker A

And so you'd give it some feedback and then it would look less terrible, but not great.

17:13

Speaker A

And the trap that this person has fallen into is not planning the piece of work before giving it to a new conversation.

17:22

Speaker A

So if we spend a lot of time planning out a piece of work, Claude is only really editing text.

17:30

Speaker A

And if we... Get called to immediately build some sort of website or some little artifact and we make mistakes, then it's more expensive to edit that than it would have been to just edit the plan before we built it.

17:44

Speaker A

So if you take anything away from this section, it is always plan anything. Complicated before starting to build it.

17:50

Speaker A

And remember, you want to use Opus for your plan so that you get a really good plan.

17:54

Speaker A

Then you want to start a new conversation, hand that plan to Claude, and you will spend far fewer tokens.

18:00

Speaker A

Claude will make far fewer mistakes and you'll probably get what you want in a shorter length of time as well.

18:06

Speaker A

Okay, the next one is about the size of the files that you're giving to Claude.

18:11

Speaker A

I want you to stop giving Claude PDFs. If you feed a huge PDF into Claude, it just burns through tokens trying to process the text, but also all of the formatting in that PDF.

18:25

Speaker A

So before you want to give a file to Claude, find an online tool that converts a PDF to a markdown file and then feed in the markdown file.

18:35

Speaker A

And you will probably use 5% to 10% of the tokens that you would have used by feeding in the PDF.

18:41

Speaker A

This also goes for screenshots. It is quite expensive from a token point of view for Claude to process visually what is in an image.

18:51

Speaker A

An image is quite a big file. Claude has to read every single pixel and then try to figure out what's in this image.

18:57

Speaker A

I've noticed that people are often taking screenshots because they haven't connected the underlying tool.

19:04

Speaker A

So for example, when I was using my CRM Attio for the first time, it wasn't connected to Claude.

19:10

Speaker A

And so sometimes I'd be taking a screenshot of the pipeline, pasting that into Claude, and then having a conversation about the state of the pipeline, which is quite expensive.

19:19

Speaker A

When I connected Attio through MCP to Claude, I could just ask a question and Claude could see the pipeline immediately and didn't have to process the image.

19:28

Speaker A

So I want you to ask yourself, Every time you paste in a screenshot of something, are you doing this because you haven't actually connected the underlying tool?

19:37

Speaker A

And if you are, then go and look for an official connector to that piece of software or to that tool, because it will use fewer tokens getting the information via MCP than it will.

19:49

Speaker A

Processing a huge screenshot. The final habit is a more advanced one, and it's available to people using Claude Cowork and Claude Code.

19:56

Speaker A

And this is using sub-agents as much as you possibly can. A sub-agent is a new conversation that happens inside your main conversation with a fresh context window.

20:09

Speaker A

And so what this means is that if you ask Claude to use sub-agents, then instead of feeding in the entire conversation so far and having all of this back and forth.

20:20

Speaker A

Claude sends a briefing to each of the sub-agents and they complete the task in a new thread with a lot less context and therefore far fewer tokens.

20:30

Speaker A

I'll give you an example. Let's say that I was going to ask Claude to look through all of my granola transcripts from the last week.

20:36

Speaker A

And this might be 20 or 30 transcripts. And let's say that I've been doing some planning around my week.

20:42

Speaker A

And so this conversation already has quite a lot of context in it. Without sub-agents, Claude would sequentially, one at a time, go through 20 transcripts and the conversation context window would just keep expanding and bloating.

20:57

Speaker A

And the number of tokens I'm using would just go up and up and up.

21:00

Speaker A

If I do that with sub-agents, not only does Claude do parallel work, so you might have four or five or sometimes 10 sub-agents working in parallel, which is much faster, but actually each sub-agent only goes to look at the Gronga transcript, summarizes it, and then sends that report back to the main thread.

21:19

Speaker A

So it's not taking in all the context of the previous planning session. It's just doing one targeted piece of work in a very token efficient way.

21:28

Speaker A

And then reporting back to the main Claude and carrying on. To use subagents is very simple.

21:33

Speaker A

Just ask Claude Cowork to use subagents, or you could edit your Claude.md file and ask Claude to use subagents where possible.

21:41

Speaker A

This is something that Claude Code does quite naturally, but in my experience, Claude Cowork needs to be asked or sort of reminded to use subagents.

21:51

Speaker A

So use subagents, keep your main conversation thread pretty clean, and you will use far fewer tokens.

21:58

Speaker A

And there's a final bonus step that a lot of us have probably forgotten, and that is that we don't need to use AI for everything.

22:05

Speaker A

Using Claude to do a web search, for example, will require quite a lot of tokens.

22:10

Speaker A

Opening Google... Finding the information and then giving that to Claude is free. Another example is asking Claude to go and look up, for example, customer's email address in your CRM or, you know, to read the notes of the latest Granola file.

22:26

Speaker A

That is going to use quite a lot of tokens. You could just open Granola and...

22:30

Speaker A

read the transcript, or you could just open the CRM and get the customer's email address.

22:36

Speaker A

So you don't need to use Claude for absolutely everything. And if you are hitting limits regularly, then I would suggest that you try to take the free action wherever you can, which might just be opening a website, getting some information and giving it to Claude

22:51

Speaker A

versus asking Claude to look things up and spend tokens on something that would have taken you usually less time.

22:58

Speaker A

So I hope you've enjoyed this video. I've packaged everything that I've mentioned into links in the description.

23:04

Speaker A

There's prompts there. There's snippets that you can paste into your code.md. There's visuals and frameworks that you can save and refer back to.

23:13

Speaker A

If you found this helpful, then do share it with someone else. Hit the subscribe button if you want to get more tutorials like this.

23:19

Speaker A

And if you're feeling really nice, you can hit the like button, which helps this video and my channel reach more people.

Topics: Claude AI token limits AI usage optimization Claude models Sonnet model Opus model effort level thinking mode token saving hacks AI workflow

Study this video

- [Flashcards from this video](https://sozai.app/tools/ai-flashcard-generator/?from=transcript&id=74108)
- [Quiz on this video](https://sozai.app/tools/quiz-generator/?from=transcript&id=74108)

Free, made by AI from this transcript. No signup.


---
This is the markdown twin of https://sozai.app/transcript/never-hit-claude-usage-limits-hacks/ — the same content, without the markup.
Published by SozAI (https://sozai.app). Reuse and quotation are allowed with attribution and a link back.
Machine-readable index: https://sozai.app/llms.txt · data API: https://sozai.app/api/