Peguei as regras vazadas do Claude Code e colei num modelo de 15 centavos (RESULTADO ABSURDO!)

Author: Matheus Battisti - Hora de Codar · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Peguei as regras vazadas do Claude Code e colei num modelo de 15 centavos (RESULTADO ABSURDO!) · Channel: Matheus Battisti - Hora de Codar · 13:59 · Watch on YouTube

Topics: AI agents, Claude Code, Prompt and context engineering, Software engineering with AI

Summary

The video explores a GitHub repository collecting leaked AI system prompts and explains how these hidden instructions can shape an agent’s responses. Rather than adding thousands of lines directly to a model’s context, the presenter demonstrates using an AI model to distill relevant instructions into a shorter Portuguese prompt. He applies a planning prompt based on OpenAI Codex to create an expense-tracker app, then uses a condensed Claude Design prompt to improve its layout and user experience. The presenter reports that the design improved, while noting that leaked prompts may not reflect the companies’ current instructions and that results should be tested.

Key points

A system prompt is an instruction the AI receives alongside the user’s message, influencing how it responds.
The video introduces a GitHub repository containing collected system prompt leaks from several AI companies and tools.
Some source prompts are very long, so copying them in full may consume context and include unnecessary material.
The presenter asks an AI to condense a source prompt into a short Portuguese version focused on a specific task.
A prompt derived from OpenAI Codex planning instructions is used to guide planning for a personal expense-tracker app.
The planning model asks clarifying questions about features such as editing and deleting expenses, categories, and filters.
A condensed prompt based on Claude Design is then used to request a phased layout, UX, and UI improvement plan.
The presenter says the app’s layout improved, while emphasizing that functional systems do not always need elaborate visual effects.
The presenter cautions that leaks may not match the companies’ current system prompts and treats the results as experiments.

Chapters

Start	Chapter	Summary
00:00	System prompts and the leak repository	The presenter explains what system prompts are, introduces a repository of collected leaks, and describes how the prompts can inform other models.
03:06	Why to condense long prompts	After showing examples in the repository, the presenter explains why using thousands of lines directly may be counterproductive.
04:38	Creating an app-planning prompt	The presenter uses a reference prompt from OpenAI Codex to generate a concise system prompt for planning a software project.
07:16	Planning the expense tracker	The model asks clarifying questions about the app’s features, then produces a plan for the personal expense tracker.
09:19	Applying Claude Design and reviewing the result	The presenter creates a condensed design prompt, requests phased improvements to the app, and reviews the updated interface.
12:27	Takeaways and closing	The presenter reflects on the value of functional design, shares where to find the repository, and closes with channel and training announcements.

Quotes

Basically, these are the instructions the agent reads before answering you. — 00:00

it’s going to clog up the context and, I think, do more harm than solve problems for us — 03:37

focused only on planning a software project without writing any code — 05:42

We can’t guarantee if this leak is the one they are still using — 08:48

a system that works and solves the problem sells better than all that beauty — 12:27

Concepts and entities

system prompt, DeepSeek Flash, OpenAI Codex, OpenCode, Anthropic

Full transcript (original language)

00:00 Guys, system prompts for Fable, ChatGPT , and Claude have been leaked and gathered in a repository so you can access and use these system prompts to improve your AI agents. Basically, these are the instructions the agent reads before answering you. In other words, even before your prompt, it’s already configured with these system prompts, and that’s how it responds based on the instructions it was given. We’re going to explore this

00:30 repository, understand how it works, and apply it in practice. Check it out. So guys, just to give you a bit of context on this system prompt issue, because some people might be a bit lost on the subject, right? It was very common back in n8n, for those who used it, to configure a system prompt for an agent. For example, if it were a receptionist agent, you’d tell it: “Be cordial, only answer questions about the company, right?”“Don’t go off-topic .” Even prompt injection issues could be handled at this stage, you see? So,

01:01 basically, it’s an instruction that the agent reads before it answers you. These are leaks that happen from time to time, and this person gathered them into a repository so you can use them in other models. So, it’s like using the Fable system prompt in DeepSeek Flash to improve its effectiveness or make it respond like Fable, right? The idea, of course, is that it will never be 100%identical, but we’ll use the Fable parameters to make DeepSeek respond that way, okay? That’s the

01:32 big idea. So, here is an explanation of what a system prompt is for you guys. We have our message here, and there’s this invisible instruction we don’t see, which is the system prompt. This is combined in the AI model; it receives what we send, plus the system prompt, and then it forms the response. That’s why I said it’s something pre-configured; we don’t see it, but they managed to leak it. It doesn’t list the exact source for everyone, but there’s a lot of information available, and it was all put into this repository I’m talking about. Well,

02:04 the repository has over 60,000 stars on GitHub, right? So it has credibility; it’s not just something random thrown up there, it has been maintained for a long time. It has many companies, so you can choose whatever you want. Some of them are here, like, you know, Claude, Gemini, Google, Core, ChatGPT, with the LLM models, for example, Grok, GitHub Copilot, and so on. So, take a look, guys, here are some examples of what you’ll find there. The rule comes

02:35 with the “why.”“Oh, don’t use X because of Y.” So there will be some prompts already placed there instructing what the AI has to use. Good and bad examples, contexts separated into blocks, and who wins when rules conflict. So everything is well organized there, and it has several categories. Let’s head over to the repository. Look, this here is the repository, guys. System prompt leaks, okay? We can see here that it was updated 19 hours ago, right? Usually, that’s when a model is released,

03:06 which is why some things are older, like two months ago. But here we can see each one of them, the leaks, the recent additions or changes, right? And basically, these are files that we can enter and understand what the prompt is , right? For example, here’s Claude 3.5 Sonnet. If I go in here, I have the system prompt, look, it’s all here, right? The big problem is that, man, we’re not going to use this entirely, right? Because if we were, man, I don’t even know how many lines are here. Look

03:37 here, 6,600. Man, it’s going to clog up the context and, I think, do more harm than solve problems for us, right? So, I use a strategy here, right? I’m starting to use it, obviously, guys, this is all testing, so I’m starting to test to see if it really has real results that will also help you get results, okay? Without using this system prompt the wrong way, alright? So, I’ll leave the link here in the description as well. Oh, this is the one, look, so you can understand better

04:08 , choose the agent you’re going to use, and so on, okay? So I’m going to do a practice run here with you so you understand how I’m testing the leaks. But first, don’t forget, guys, if you want to learn AI-assisted coding, that is, creating projects with AI, there’s my super complete training here, with over 30 hours of course material, okay? I’ll leave the link in the description, and you can watch this video where I explain the entire training. And of course, you can check the curriculum, see the projects you’ll create;

04:38 everything is detailed here regarding what you’ll learn and what you’ll build with me, alright? Lifetime access, student group included, a super complete training for you to learn programming with AI and build any kind of system. Deal? So let’s head over to VS Code, where what I did here, see, I opened the OpenCode interpreter with flash. So, obviously, most of you will use it as a standard free model, and create a system prompt or a file

05:10 derived from the system prompt. There are the instructions from OpenAI, Anthropic, or whatever model you want to use, okay? So, it works in the following way. I first give an instruction prompt, okay? Which will do the following: read the prompt of an AI tool at this link. So here, for example , I have a link for the OpenAI Codex planning mode, okay? So man, it’s an OpenAI system prompt for system planning. What am I going to do? Create

05:42 a new system prompt from it in Portuguese with a maximum of 300 words, focused only on planning a software project without writing any code, asking only about decisions that really change the plan and delivering a clear, verifiable plan. Rules. Take only the instructions that serve this focus. Remove mentions of tools, use short sentences, answer only with the final prompt. So, what did I do here? I told it to create a prompt that will help me plan systems. You can do this by asking it to output it to a markdown file. You

06:12 can do it as a prompt, like I did here. You can just send the reference link and send your prompt, okay? So there are several ways to do this. I like this because I can read the prompt. So it sends me the output here and I can copy it as a system prompt, okay? Look, it gave it to me here. So there is the whole system prompt it generated. Of course, you can improve it, you can modify my prompt, but it’s an idea of how you can do it. So here I’m going to copy it, okay? And I will paste this

06:44 system prompt here. I will send my prompt. So the AI will read the system prompt, read my prompt, and form the final answer based on those two pieces of information. It’s as if we were adding a new system prompt, right, which will be included in the context, but I’m not including those 6,000 lines , or however many there are in the link I put here. I’m filtering it so there isn’t too much unnecessary stuff, you know? For example, if we take the Cloud Design plan here, their own design system, man, it has 2,600 lines, this

07:16 could hinder more than it helps. So, what I’m doing is this preliminary filter so that the information is more condensed and I can still get value, right, when using this guy. I’ll press enter here. And then it sent the system prompt as I requested and my own prompt . So, I want to create a personal expense tracker app. It needs to have a form to add expenses with description, amount, category, and date, a list of expenses, and the total by category, saving the data in the browser, okay? Look how interesting that was. It already asked questions because some

07:46 information was missing, purposefully, obviously, but it’s already acting as the system prompt requested, asking about what was vague. So I want, I’ll choose the recommended option here. Then here, can the expenses be edited or deleted? Yes, edit and delete. How to handle the categories? A fixed list in the code or customizable by the user . I’ll want, then do you need to filter or group by period? It’s here, it needs a filter. And then I trust it, and that’s it. Oh, if it has any more questions, it will ask, otherwise, it will follow the rest of the planning and my execution here, just to show you

08:18 guys as well. While it creates it, I’ll show the plan mode I used here. So, look, this was the file it read, okay? This is what I used as a base to do the project for this video. So it’s the Open Codex plan mode, which is likely when we open the codex there and activate the plan, it opens a system prompt, see? So it delivered the plan to me. Then we have to read it, obviously, right? And ask it to build if we want to or change something. I’ll ask it to create. We can use, for

08:48 example, Cloud Design to perhaps improve the product’s design and others there so we can make an average AI get better results, okay? That’s the main idea of the leak and, of course, to understand how companies work behind the scenes. We can’t guarantee if this leak is the one they are still using, right? Because I believe this also gets updated, right? System prompts occasionally, but we have an idea of how companies act. We can implement this in our systems too. We can use

09:19 those system prompts we have at OpenAI to plan our AI-embedded systems to guide an agent. It’s a good way to use it too, let’s put it that way, okay? So here it created it, guys, this is DeepSeek Flash. It really is very fast. I’m going to open the example project here, right? Of course, just something fictional, right? For you to see. It created it here and so on. It has all the requirements we asked for, clarified, with what we had there in the system prompt additions for those doubts that weren’t clear to DeepSeek.

09:53 I’m going to make that proposal I mentioned to you before. Let’s try to add Cloud Design now into the mix, right? Let’s see if it can improve the application’s design. You have a base here. This is super standard, right? Nothing more than a basic CSS framework . And here I will clear it and put how I would use Cloud Design here, okay? So I’ll ask it to read it again, right? So read the system prompt in an AI tool at this link. So Cloud Design, okay?

10:23 Create a new system prompt in Portuguese from it. Same thing with the rules here, okay? Let’s wait for it to send the prompt. It sent it here, we copy it. I’ll put it like this, look. I have a personal expense tracking system here. The design is pretty bad. I want to improve it. And I’ll put here: check the project and help me with an

10:54 implementation and improvement plan for layout, UX, and UI. Okay? Something like that. I sent the system prompt, I sent my request, and here it will read the project and give me an improvement idea. Oh, once again this one also has questions, right? This one is from Anthropic, it’s from another company, right? It’s not from OpenAI. And so I’ll want a modern financial look here. Let’s work with this style here, okay? To really make it stand out, because

11:24 the minimalist one I think will be too similar to what we have. And here it suggested doing it in phases, okay? I will put here: apply all the improvement phases in sequence. Just so we can take a look here, right? I like to do things step-by-step, but these are more practical examples, so let’s see if there will be a real impact on our design using this Claude design system prompt, right? Hey guys, so it finished the six phases, that was a lot . It froze in the middle there, the

11:56 opener session out of nowhere, but that’s okay. Let’s see if we can make a comparison here, right? Unfortunately, it already updated here, you know? It didn’t keep the previous view, but it did add some features here . It improved the layout quite a bit, as we can see here. So everything looks better for a financial control system. It also added some warning modals and things like that. So, the layout really

12:27 did turn out much better. Of course, it doesn’t have that cinematic stuff everyone is selling right now, making those 3D sites that move with the scroll, but man, a system isn’t always about that, okay? In my own experience, a system that works and solves the problem sells better than all that beauty that we’re not sure even works in the end, right guys? Well, so look, I’ll leave the repository here in the video description for you to access and

12:58 see if it’s worth it for you, okay? I’ll test more for further analysis, if you want, leave it in the comments as well. Vibe Code training, if you want to learn several tools together, like Claude Code, Codex, Antigravity and others, creating various projects too, working with various business rules and different integrations. This is the training, okay? I’ll leave the link in the description so you can access it and check it out. There’s a video here where I explain everything as I said, but if you have any questions, just message us on WhatsApp.

13:28 All right, guys? Over to you now. Leave a like, and subscribe if possible, okay ? Also, don’t forget, we have weekly members-only videos, alright? So, every week I release a video for the members, okay? So, become a member. There’s a VIP group on WhatsApp and I’m doing weekly live streams here too, okay? There isn’t a fixed day for them. I’m doing it whenever I can actually manage to, but usually in the afternoon, around 3 o’clock, bringing up trending

13:59 topics from AI week, so to speak, and testing some models live so we can see how they behave, with tests where I can show errors and such. Something more flexible than edited videos like this one, okay? So, I look forward to seeing you there and in the next video, thanks .