**Eu Resolví o Desperdício de Token do Claude Code (Apenas Copie) — Transcript & Summary | SozAI**
Source: https://sozai.app/transcript/resolvi-desperdicio-token-claude-code/

Learn how to reduce token waste and improve Claude Code usage with audits, habits, and model selection tips.

## Key Takeaways

- Token waste in Claude can be audited and reduced with a personalized prompt.
- Removing unused connectors and managing sessions carefully saves millions of tokens.
- Proper session management habits, like using 'clear' and avoiding fragmented messages, improve efficiency.
- Choosing the appropriate Claude model and effort level optimizes cost and output quality.
- These strategies help users avoid hitting usage limits and get better results from Claude.

## What the video covers

- Claude has been using more tokens than usual, causing faster usage limits and poorer model responses.
- Felipe developed a personalized audit prompt to identify individual token consumption and optimize usage.
- Opening new conversations in Claude can consume a large number of tokens due to preload costs.
- Removing unused connectors and MCPs drastically reduces token consumption per session.
- Using 'clear' between tasks is recommended over 'compact' to avoid token waste and maintain context.
- Avoid sending fragmented messages and use a single model per session to optimize token usage.
- Editing prompts instead of correcting them saves tokens by erasing wrong inputs and answers.
- Adjusting model effort levels affects output length and token cost; lower effort can save tokens.
- Use the right Claude model for each task: Haiku for simple tasks, Opus for building, Fable for architecture and review.
- Following these steps can significantly extend token limits and improve response quality.

## Chapters

1. 00:00 Introduction to Claude Token Issues
2. 01:31 Personalized Token Audit Prompt
3. 03:01 Understanding Token Preload Costs
4. 04:31 Removing Unused Connectors to Save Tokens
5. 05:45 Using Clear Between Tasks
6. 07:16 Avoid Fragmented Messages and Use Single Model
7. 09:52 Editing vs Correcting Prompts
8. 12:14 Adjusting Model Effort and Choosing the Right Model

Answers

## Questions about this video

Why is Claude using more tokens than before?

Claude's token consumption has increased due to factors like longer sessions, high effort levels, and preload tokens per conversation, leading to faster usage limits.

How can I find out what is consuming my tokens in Claude?

You can run a personalized token audit using Felipe's prompt available in the free community, which analyzes your individual token usage and provides a detailed report.

What are the best habits to reduce token waste in Claude Code?

Use 'clear' between tasks instead of 'compact', avoid fragmented messages, use a single model per session, and remove unused connectors to save tokens effectively.

## Full Transcript — Download SRT & Markdown

00:01

Speaker A

Recently, people all around the world have been saying that Claude is using more tokens than usual. These complaints have always happened, but now it seems it has truly reached a very bad level. And that is not all.

00:13

Speaker A

Besides tokens running out faster, they are also saying that the models are returning worse results, as if the models were getting dumber. But I found a new system within Claude that has more than tripled the amount of usage I

00:25

Speaker A

have available on my models. And in this video, I will show you exactly how to never hit your usage limit in Claude again and also get better answers from the models you already use today. So grab your coffee and let's get straight

00:38

Speaker A

to the point. And the first thing I want to tell you is that it is not your imagination that Claude is spending more and getting dumber. This is actually happening. We have several reports about this here, many people

00:48

Speaker A

complaining about it, reporting, for example, look at this, the Max 5X plan took 5 hours of limit and used it up in 23 minutes. Three simple prompts blew through the entire session, meaning it is really happening, we are having a

01:00

Speaker A

higher token spend, yes. Therefore, we need to be more careful. And this will get worse, because Anthropic will cut 17% of the weekly limit we currently have in Claude. So, if the limit is already low, it will get even smaller.

01:11

Speaker A

And this announcement is being made to take effect on September 14th. In other words, if you are watching this video before, start preparing. If you are watching it after, this could be the salvation for your tokens that are

01:21

Speaker A

being spent much faster. And starting here, then, at our first level, we will have the audit. To know how to improve your token spend, you need to know what is consuming your tokens. For that, we are going to run a full audit. And here

01:31

Speaker A

, it is individual. This is the interesting part, because what is consuming Felipe's tokens is not necessarily what is consuming your tokens. And that is why I developed this prompt that I ran, I did my audit, and it more than doubled the limits I

01:44

Speaker A

had for tokens to use. So this here is individual. There are many things we will see here throughout the next levels that apply to everyone, but this specific part you have to run on your computer; you need to understand what

01:56

Speaker A

you're doing wrong or what you can do better. To do this, you can just copy this prompt of ours that will be available in our free community. Just find the link in the description or in the first pinned comment. You will copy

02:07

Speaker A

this and pass it to your Claude code. So you will open the tool, paste it, send it, and from there it will generate this document for you, this complete audit. And the result you will get after this audit will be this token

02:18

Speaker A

dossier, where it will explain to you detail by detail where you are spending and what you can do to improve. So here at the beginning, we will have it explaining exactly where every single token we use is going, including

02:30

Speaker A

projects and the cost of using multiple sessions. It will explain here in detail. It even shows that look, I spent a very reasonable amount of tokens here. So it's giving me all this in detail, but if we come here, it will

02:42

Speaker A

list the diagnosis with seven problems in order of impact according to what Felipe does. So here we have sessions that are too long without using clear.

02:49

Speaker A

We will see this in a little bit. This here is very important. Preload at 74,000 per session. And so, just for you to understand what this preload is, if we enter our Claude here and click on context window, it will show exactly

03:01

Speaker A

where our tokens are going. So here I, Felipe, when I open a conversation without having said anything to my model, I have already spent around 70,000 tokens just by opening a conversation. And each new conversation I open, those are 70,000 tokens that

03:14

Speaker A

are being burned right off the bat because of this. It will give a full breakdown of where this token burn is going and coming from. So it will be here talking about the messages; in this message I spent 28, system

03:24

Speaker A

messages 17. So it will list all of this, and our dossier will also have this same analysis. In other words, when it talks about preload, it means things I spend in every session. Then it will have output inflated by effort

03:36

Speaker A

set to high. Effort is the model's effort level. So, imagine you are using GPT-4o and you set it to maximum effort. This will make the model respond to you, return outputs, that is, larger, longer responses, which don't

03:48

Speaker A

necessarily help you, aren't necessarily better, but definitely spend more tokens. The tokens for this model on X High are more expensive, and besides that, it will return a larger output that will consume even more tokens. So this is terrible, and notice

04:00

Speaker A

that these are things that even happened with my own usage. But having said that, it will go through several things, and down here it will have tips, a guide on how to better use Claude Code, and it will give you some of

04:09

Speaker A

these tips based on your usage. That's why it's important for you to run this audit to understand your needs. And once you've run this, there's another thing I want you to do right at level one, which is cutting out unused MCPs

04:20

Speaker A

and connectors. To do this, you'll come here to your Claude Code, go to connectors, manage connectors, and you'll look at the list of connectors you have, of tools you have within your Claude Code, and delete those you don't

04:31

Speaker A

use. Mine is small here, as you can see, because I've already done this. This was responsible for going from a window, a preload, just from opening the window, of 170 to 180,000 tokens, to today, to below 70,000. So, I reduced

04:43

Speaker A

more than double the amount of tokens I was using just to open the window. This isn't my entire token-saving process.

04:48

Speaker A

There were other things I did too, but this alone solved a lot for me, because imagine that today I spend 100,000 fewer tokens per window I open every single day in my Claude Code. Imagine how much that adds up to in a week, a

04:58

Speaker A

month, a year. I'm talking about millions, maybe even trillions of tokens I'm avoiding spending, depending on how much I use this tool and open new conversations. So come here without mercy and remove everything you don't use. "Felipe, but what if I remove it

05:09

Speaker A

and need to use it again?" Later, you just come back here and connect it again. That is the wonder of our Claude Code. It is very easy for us to make these links and connections. So if you're not using it, you can cut it

05:19

Speaker A

loose. When in doubt, cut it loose. If you need it, you just connect it again later. With that, we can then reach our level two, which is habits. These are universal habits that will work for me, for you, and for everyone watching this

05:30

Speaker A

video. And here, more than habits, these are some little hacks that I myself used to get wrong while using Cloud Code. So it doesn't matter if you have experience or not. The fact is that we often end up making mistakes.

05:41

Speaker A

The first one, and I'll make it very clear, is to use clear between tasks.

05:45

Speaker A

What does that mean? Imagine you're talking to your Cloud Code about pigeons, and you said: "Look, I love pigeons, they fly, pigeons are in the city, or pigeons are wonderful, I like feeding pigeons." You talked about that. Then you want to talk about dolphins,

05:56

Speaker A

what are you going to do? You're going to run clear. Felipe, but couldn't I run compact too? Yes. And no, you can, but you shouldn't. Why? Compact will take everything you talked about regarding pigeons; it will—look at all the work—reprocess everything, it

06:09

Speaker A

will read everything again. Therefore, this will spend tokens, it will spend time to store in its information, eventually, something important you might have discussed, in case you want to return to that subject later. But that sounds very cool, depending on the

06:20

Speaker A

project you're talking about. However, it doesn't work. It doesn't work. It is very bad. It doesn't keep the context the way it should. In short, it ends up having a very poor result. You won't be able to keep the context.

06:32

Speaker A

need, so it's not worth using. So it's better for you to use clear. Clear will wipe everything. It won't even pretend to keep any context, like Compact does, but it will free up your context window . With your context window cleared,

06:43

Speaker A

your model becomes smarter, because we have a context window limit there. So, up to 200,000 tokens, 400,000 tokens, we get very good results when exchanging ideas with the model. When it goes beyond that, the model starts to get dumber—that's a fact—because

06:56

Speaker A

it starts losing context. But here, if we run clear, it will reset the context window, meaning the model will have a clean slate to hear what you have to say, will give you better results, and will spend fewer tokens. You'll spend

07:05

Speaker A

fewer tokens because for the responses you send and it sends back, it won't have to reprocess all that context since you cleared it, there's no more context. And the advantage of using clear, therefore, is that you won't have to reprocess all that information

07:16

Speaker A

like Compact does, which ends up spending tokens only for you to have no practical result from it. Second thing, use a single model per session. This is where 90%of people go wrong, because imagine I started talking to my model

07:27

Speaker A

about a subject and I'm using, for example, Fable 5. So I'm here all excited, talking, and I'm going to ask it something simple. Hi, please, I would like you to spell joy. I send it any nonsense, the simplest things, the

07:39

Speaker A

simplest questions possible. And I sent that using Fable 5. Imagine I want to keep spelling several words, but I know I don't need Fable 5 for that. And what do many people do? Many people come here and switch the model. They simply

07:50

Speaker A

come here and switch to Sonet 5, for example, or Haiku, which Haiku handles this job easily. But there is a problem . When you're doing this on a project and you say: "Gee, from this point on in the project, it's simpler, I'll use

08:02

Speaker A

Haiku here or Sonet." When you do this, when you change the model, it has to reprocess all of that. So the new model will have to reprocess everything you talked about to store it in the memory we call cache memory, which is a little

08:11

Speaker A

memory that stays there in the model so it understands what's happening. The Fable 5 model did its whole process up to there, it has this memory, it was the one that did it. Now, when we change models, the new model has to

08:21

Speaker A

relearn everything all over again. Guess what? This spends tokens. This spends a lot of tokens. So I'm telling you, you don't save anything by doing this, nothing. On the contrary, you might even spend more thinking you're saving. So don't do that. Stick with a

08:33

Speaker A

single model until the end of your task . Want to do a new task? Run a clear, start a new task with a new model, then it works. Then it's good. Another mistake that many people make, by the

08:41

Speaker A

way, many people make this mistake on WhatsApp, it's frustrating, wanting to keep sending several messages. So, instead of sending a single message with what they want, they send: "Hi, how are you? How are things? Can you help me with something?" So, they send

08:51

Speaker A

four or five messages to say something that could have been sent in just one.

08:54

Speaker A

This isn't just annoying on WhatsApp. This is also annoying in Claude. It also costs you money in the cloud when you do this. So, don't send five separate messages. The conversation gets reread five times. Because remember, every time you send something

09:05

Speaker A

to the model, it rereads everything you've talked about. Every time it gives you a response, it rereads what you've discussed. So, if you can send less, you'll have a better context window. In other words, you get the most out of the model's intelligence,

09:15

Speaker A

faster, because it has almost nothing to reread; and it will deliver something cheaper, because it will be more efficient, use fewer tokens to read what you want, and provide a better result faster. So, group your questions, especially when you have

09:27

Speaker A

messages like I do. For example, I click here in my Claude and I just start talking, and everything I say gets transcribed and passed into the prompt. So, look, I just did it, let go , and that's what it did. This makes

09:37

Speaker A

things much easier. So, when I want to get something done, I prefer to explain it very well, write a prompt that's even quite long, but a single complete prompt with everything I want and everything I've thought of, at least so

09:47

Speaker A

it can try to solve it in one or two prompts at most; because then I know I'm using the full power of Claude, while paying as little as possible.

09:52

Speaker A

Therefore, don't send fragmented messages, never, if possible. Now we arrive at topic four, edit, don't correct. What do I mean by that? So, imagine we're here in Claude and I ask it a question. How old is the oldest

10:05

Speaker A

person in the world? I asked this question here, and sent it. And now it's going to answer me. So it's going to give me the answer here. I actually love using Haiku 3.5 for this. It's a super cheap model, great for us to play

10:15

Speaker A

around with, you know, asking questions . Obviously not, that's sarcasm, but anyway. It's giving it to me here. Good question. Look at how it’s already going to wander off on a simple answer it could give me. A simpler model would

10:23

Speaker A

give me a simpler answer. But we’ll get there. Imagine then what it's giving me, but I say: "Wow, I don't want to know the oldest person, I want to know the oldest woman." So what would I do? I could copy this here,

10:34

Speaker A

paste it, and instead of putting person , put oldest woman here. Then I come in , swap it for woman, and send it again.

10:40

Speaker A

Doing this is horrible because it will have to reprocess everything all over again. What is the little hack and the tip here for you to get the answer you wanted? You are going to come here and edit this message. So you’re going to

10:49

Speaker A

come here, look, go back to this point. You go here, look, it erased that context because it was wrong. You want to rephrase it, and it will give you the opportunity to ask again. Then you come here and change person to woman,

10:59

Speaker A

and then you do your search correctly. In other words, don't leave the error, the wrong answer, and then ask the correct question. Fix the error, come here to this little circle, and from here you click go back to this point.

11:11

Speaker A

It will erase everything that was done wrong. The question that was asked wrong, and therefore, the answer that was given wrong. This way you also save a lot of tokens. This is a golden tip that helps a lot. No kidding. This

11:21

Speaker A

helps a lot when you need to correct some information, especially larger information when you're building bigger projects, more professional projects.

11:27

Speaker A

Finally, we’ve reached our fifth hack , which is to send text instead of PDFs and screenshots. Screenshots. When you send a screenshot of something, it is extremely costly for the model. The model spends too much when you send a

11:38

Speaker A

screenshot. And yes, I am the biggest culprit of this. I raise my hand, I send a ton of screenshots to my model.

11:43

Speaker A

I’m stopping that. I’m trying to rephrase, I’m trying many things that I could explain with words and that I used to simply send as screenshots.

11:48

Speaker A

I’m trying to rephrase this so that I can indeed explain as much as possible.

11:51

Speaker A

Eventually, we will have to send screenshots, yes. It’s part of the process, it helps the model understand what you want to create, but we have to use it sparingly, with great calm, screenshots and PDFs. So, here's the tip: you have a PDF there that you want

12:04

Speaker A

to send, copy the text and paste the text. Don't just send the PDF and ask him to read it; that will already help a lot. So, these are five little hacks that will totally change the way you use Claude, you'll have much better

12:14

Speaker A

results, more organized, faster, and much cheaper. Believe me, this makes a huge difference. But now, before I show you the third step, I want to take this moment in the lesson to invite you, because by now you've probably realized

12:26

Speaker A

that using Claude professionally isn't just opening the app and using it without any criteria. To do it the right way, we need process, reference, and method. And that is exactly what I teach inside AICloud. In AICloud, you'll find a complete track to learn

12:40

Speaker A

how to use Claude professionally and create projects that you can actually sell to clients. Inside, I open the black box of my creation process, the skills I use, templates, workflows, security precautions, and delivery standards that I apply to my own

12:53

Speaker A

clients. So, if you want to go further and start using Claude to build real things, serious things, and make money with it, I advise you to point your phone camera at the QR code appearing on the screen and come check out

13:05

Speaker A

AICloud. I guarantee you will be very welcome. Now, message delivered, let's go back to see our level three of using Claude. Level three is something I see very few people paying attention to, which is asking Claude for answers

13:17

Speaker A

without the fluff. Why? Have you noticed that when you're messaging Claude, sometimes you ask such a simple question and it sends a paragraph, two, or three, explaining things in a way that's too complex, much more than you actually need? Yeah, that happens to

13:30

Speaker A

you, it happens to me, it happens to everyone, and it's part of the model's dumbing down. Beyond the dumbing down, besides wasting mental energy reading a response that could be much shorter and simpler, it also spends tokens, because

13:42

Speaker A

the output, which is what the model returns, comes out larger, so it ends up being more expensive. So, when we ask it to answer without fluff, it involves three major things. First, asking it to do so. I'll also leave a

13:54

Speaker A

little prompt here for you to copy and send whenever you see it's sending messages that are too long, which also relates to how you use the model. So, if you've selected, for example, Claude 3.5 and it's on high, at maximum effort

14:05

Speaker A

, just know this right from the start. The higher the effort, the longer the output will be, and the more expensive that output will also be. So, if you're going to ask simple questions, use a simple model at a medium, reasonable

14:15

Speaker A

effort level. That way, you'll get the best out of the model, the best model for each task. We'll talk more about this later, but this also helps a lot.

14:21

Speaker A

Furthermore, one thing that is important for you to do is to provide style instructions. You'll copy this, send it to your Claude, and say: "Look, always follow this." It will add this to your Claude.md, and it will start

14:31

Speaker A

responding with more clarity, with shorter answers that are easier for you to read—better for you, better for Claude, and much cheaper. That is the most important part. Long-winded responses are expensive, so we need to set a communication style. Actually,

14:44

Speaker A

what I recommend is that you find your own communication style with it. You'll send this prompt here. This prompt is universal; everyone should send it. And later, when you see it's responding in a human way, the way you like, the way

14:55

Speaker A

you say, "Wow, this is cool to talk to, " tell it: "Look, please take this communication style we have now and turn it into a skill." Then you can call it the human skill. And whenever you trigger this skill, it will always

15:05

Speaker A

communicate with you in this shorter, more direct way, without too much fluff . So this will help you a lot with your results, the quality you get from the model, and, above all, help you save a lot of money. Now we've reached level

15:15

Speaker A

four. Level four is basically knowing how to use the right model for each task. So, Haiku, which is Claude's simplest model, is very good for renaming files, formatting files, asking where a file is, and other much simpler, totally manual tasks, let's

15:30

Speaker A

call them. Use Haiku; it can handle it, it can access your documents, it will explain exactly where things are, and everything you need, it helps you immensely. He is what we call a construction worker. Then we also have

15:40

Speaker A

Sonnet. Sonnet, in my opinion, is the model I like best because it's very versatile, it's very intelligent, it's very fast, so this here is wonderful and costs very little. Sonnet is really well-priced, it's one of the best value-for-money options. And you will

15:51

Speaker A

use this one when you need to write or execute code in your daily tasks. So, for example, you did something simple, a small HTML presentation: "Make this HTML presentation for me, open it in my browser." It will handle it. "Change

16:00

Speaker A

this for me, change the title, change the font." Damn, it's wonderful for building, for correcting what you are doing. Now, when you have larger, more complex projects, you will also need more complex models. And that is where we arrive at what we call frontier

16:13

Speaker A

models, where above them is just the sky, there is nothing above them, they are the best available. And we are currently talking about Opus, but we are talking about the Opus family of models and also the Claude 3.5 Sonnet

16:22

Speaker A

family, wait, the Opus and the Fable models, which are truly the best of all . So we will have Opus here as a driver . It will be very good at driving projects and long sessions, daily decisions with large context. Need to

16:33

Speaker A

build an entire project, something big, really large and robust. You will use Opus. You won't do that in Sonnet. Need to build a quick HTML page for a presentation, use Sonnet. Need to build an entire CRM with many details and

16:44

Speaker A

complexity, use Opus. And Fable comes in as that architect role. It will handle all the architecture, planning, and final review. So what do I usually do? I use Opus to build my applications . So imagine I created a CRM. I will

16:56

Speaker A

use Opus for that, without a doubt. I don't even use Fable, I use Opus. And then I leave Fable to fix the architecture, fix bugs, find bugs, perform security reviews, because Fable actually comes from the Mythos family, which is the model used specifically

17:09

Speaker A

for cybersecurity. So when we talk about Fable, we're talking about reviewing processes, reviewing architecture, not just cybersecurity to prevent hacks, but also to prevent your site from going down or having structural issues that could later affect your app. So, in short, Fable

17:23

Speaker A

works very well with that. And of course, to switch between these models, we'll use the slash model command. That part is very simple, we know. Just avoid switching these models, as I mentioned before, during the same task.

17:34

Speaker A

Execute the entire task, get everything done. Shoot, I want to change models. Start a new conversation. But more than just starting a new conversation, use the clear command, because leaving conversations open can often eat up your context. So, want to change models

17:47

Speaker A

? End that conversation, run clear, open a new chat, and then switch your model. It’s the best thing you can do when it comes to saving money. Now here comes a second tip, which is the cheap second opinion. So imagine you have a

17:58

Speaker A

very critical task; you did it with Opus, you reviewed it with Fable, and you need to ensure there are no errors at all. What I like to do in this case is come here and ask Codex to review it

18:07

Speaker A

, because I have the monthly OpenAI subscription. If you have anything from the simpler ones, like the Plus plan for 20 a month, you can already make this same request. So imagine there's the entire build of a CRM here. I

18:18

Speaker A

already built it with Opus, asked Opus to review it, asked Fable to review it, but I want to be sure, I want to see if there's any little error that wasn't caught by the Anthropic models. What do I do? I come here and make the

18:27

Speaker A

following request. Please, review the entire process we did using the Codex CLI. I send that to him. What is he going to do from here? The Claude model will call Codex. It will ask Codex, meaning OpenAI's models, to look at the

18:40

Speaker A

project. Please, OpenAI model, GPT-4, take a look at my project here and see if you can find any bugs. And believe me, at least 30%of the time for me— that is, three times out of 10—errors appear that Claude missed, whether in

18:54

Speaker A

Opus or Fable. So use this; it’s very interesting and you will spend very little to do it. Instead of opening several sub-agents to search for every detail using Fable 5, you can use an OpenAI account you already have, if

19:07

Speaker A

applicable, to perform the reviews. The same applies to Gemini and everything else. For my final reviewer, I like to use the Codex in the 5.6 Sonnet models; it’s an expensive model, but for reviewing, it’s excellent, very useful, cost-effective, and believe me,

19:22

Speaker A

it finds bugs that Anthropic models often miss. Finally, a great building tip is that your codebase—your code, your entire process, your whole project —should be read like a map, not line by line. Because imagine you want to find information and Claude doesn't

19:36

Speaker A

know where to look; the model doesn't know where to search, so it has to read the entire process just to find a single line to give you an answer.

19:43

Speaker A

Don't do that. Do what we call an "init ," which is asking to create a " claude.md" file in each of your projects. Look, when we are talking about design, you will access this folder. When we are talking about

19:53

Speaker A

architecture, you will access this folder. This will be inside a claude.md file, which is the tool, the document that the model must read before returning an answer. So, when you ask the model a question, it will have to

20:02

Speaker A

look at this document, and in this document, there will be a map of where it can find each piece of information.

20:08

Speaker A

This is the pinnacle of good organization, and it will bring you a lot of speed, even more than the savings or the quality of the response it also provides. The speed you gain from this will help you significantly

20:18

Speaker A

in your projects. Now we've reached level five, which includes some tips for you to add the icing on the cake to your projects. First, approved style becomes a skill. Whenever you like something, whether it's text or design, turn it into a skill, because then you

20:33

Speaker A

save all those prompts you spent to reach that result you liked. From now on, you will have a skill. This skill will state and explain what you want in a much more succinct, efficient, and better way than sending multiple

20:44

Speaker A

prompts. So, if you liked something, create a skill. Ah, but I have a lot of skills I don't use. Delete the skills you don't use and keep the good ones you use all the time. Because this will also help you, here's an extra tip:

20:54

Speaker A

don't keep saving all your skills, because that will also waste tokens for no reason. So, keep about half a dozen skills that you use a lot based on things you've created and think have interesting results, instead of having

21:04

Speaker A

to recreate them and spend more time and money. Next, screenshots, print screens, well, they are one of the most expensive items in the conversation.

21:11

Speaker A

Don't send those. Try to explain whenever you can, try to dictate something whenever you can, mainly by downloading a voice transcription app.

21:18

Speaker A

I use Whisper Flow for those who ask, but there is also Handy. It's free; you can use this type of service for free as well. Then here, the winner becomes a small spec. So when you like a color,

21:30

Speaker A

font, spacing, or component, you're going to ask it to create a design system. Ask it, "Look, take this and create a design system so we can reuse it next time." In other words, you won't have to keep making corrections,

21:40

Speaker A

changing fonts, changing colors, because that will already be in the design system. Finally, a critique loop . When you're building more complex projects, sometimes you'll set up a loop, which is sort of what I just explained in level four. I ask Codex to

21:51

Speaker A

do it, but first I run a round on our Claude. So I ask it to keep running the critique loop. So look, analyze the project, have one analyze the design, have another analyze the architecture and functionality to see if it's in

22:02

Speaker A

line with what we agreed upon for this project at the beginning. They will run this loop. It's not cheap, it's not fast, but they will help you a lot, because if they find errors there, they already fix them, and they don't turn

22:13

Speaker A

into errors that are much harder to find later when a client returns a project to you saying, "Hey, this is giving an error, it has a bug." It's important that we do this. And here's my tip: if you have subscriptions to

22:25

Speaker A

others, whether it's Gemini or OpenAI, use them for the final loop as well to send them a final review, because something might slip through. Things will always slip through with humans; even a full developer team misses things, imagine with AI. So it's very

22:37

Speaker A

common for this to happen. And based on that, our ladder, our levels, are ready . You'll need to audit, change your habits to good ones, ask for straight-to-the-point dialogue from the cloud, route work to the right models for each task, and focus on

22:53

Speaker A

cost-effective design. If you like something, turn it into a rule so you can reuse it. Now I want to know what you thought; did you like the video? If you did, please leave a like and especially share it, because it really

23:05

Speaker A

helps get this video out to more people . Also, I want to reinforce the invitation for you to check out AICO, join our community, ask me questions directly, and get access to high-level knowledge about Cloud Code—how to use

23:18

Speaker A

it professionally if you want to make a living from it, use it in your business , or even land clients. That's it, and now I'll leave another video on the screen for you to keep learning about artificial intelligence.

Topics: Claude AI token optimization Claude Code AI model usage token audit AI prompt engineering model efficiency Anthropic Claude token saving tips AI usage limits

Study this video

- [Flashcards from this video](https://sozai.app/tools/ai-flashcard-generator/?from=transcript&id=79834)
- [Quiz on this video](https://sozai.app/tools/quiz-generator/?from=transcript&id=79834)

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


---
This is the markdown twin of https://sozai.app/transcript/resolvi-desperdicio-token-claude-code/ — 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/