YouTube Deep Summary

Something crazy just happened. Someone leaked the system prompts models and tools of a lot of different AI platforms. Cursor, Windsurf, Replet, Manis, even Lovable and Devon. There is some really surprising information about how these tools actually work. Along with them, some open- source tools like Bolt and Rue Code were also leaked. Their system prompts were revealed too, and these prompts are what tell the AI agents how to function. I went through most of the leaked system prompts and found some pretty amazing things. That is what I am going to share with you in this video. This is going to be a pretty interesting video because of the leaks and the information we have gotten from them. The first thing I want to tell you is that many people still do not know which tool is better, Cursor or Windsurf. After reading the system prompts, we finally have an answer. The clear winner is Windsurf. Here are the clear reasons why Windsour's Cascade outperforms Cursor. The first reason is that it has a richer tool set with a lot of built-in apps, while Cursor only offers basic apps that force you to leave the IDE again and again. Cascade offers some crazy tools. One of them is live browser preview, which means whatever you build that needs to run on a browser runs directly inside Windsurf. You can reference things directly in there, which makes coding front-end apps much faster. Another feature they introduced is one-click deployment where they have partnered with Netleifi to make deployment really fast. They also have advanced project navigation with better search methods that cursor lacks. Their file system browser is also far better than cursor. All of these tools create a richer tool set that makes Windinsurf much better to work with. Cascade also has persistent memory which means it remembers what you told it while cursor resets memory every time you open a project. Cascade has memory crud. It can create memory and store things like user preferences, code snippets, API keys, and project milestones, which helps you track your project's progress. It also has automatic retrieval, so when Cascade restarts, it brings back relevant memories based on semantic matches. You keep working in the same way even after long breaks. Even if you return to your project after a week, Cascade will remember. Cursor lacks all of these features. Another reason why Cascade is better is because of asynchronous orchestration in the way it handles tasks. Cascade can run commands while it is writing code and editing files. Cursor follows a synchronous process. It edits something, runs tests, waits, and edits again step by step. Cascade handles everything in the background. It also checks the status of running processes automatically while cursor does not even support background tasks. You can multitask with Cascade which you cannot do with cursor. This was all extracted from the leaked system prompts so it is authentic information. Cascade definitely outperforms cursor by a solid margin. We also have some takeaways from the leaked manis agent prompt. The first is that it offers full sandbox control. Manis boots up a real Ubuntu environment with pseudo rights. It can install different packages like Python or Node packages, spin up its own servers, and even expose its ports to the internet. When we watch Manis code or perform tasks, we are seeing a full virtual engineer that has its own space and can do anything within it. Manis runs on a six layer agent loop architecture. First, the user sends a message. Manis analyzes the message, selects the tools needed to perform actions, observes the result, plans again, and picks another tool if needed. All of this happens while its internal knowledge is being updated. This structure allows Manis to function like a real world human engineer. Manis treats the browser as a first class tool. It does not use Selenium to automate tasks. Instead, it has native functions for clicking, typing, scrolling, and even executing JavaScript code. It interacts with the browser the same way a human would. Manis is also modular, which helps reduce hallucinations and makes its results more reliable. In the leaked files, there was a deploy and exposed port feature. This means Manis can push static sites or even Nex.js apps live. It can also clone Git repositories and install dependencies, making it a true all-in-one automation engineer. Manis has a built-in prompting guide as well. If it gets stuck at any point, it either asks for input from the user or reprompts itself automatically. Guardrails are also set in the system prompts, making sure that it only works inside its sandbox and does not use any external accounts. I will not be able to cover everything here. But if you want to explore other tools like the open-source ones, there is an easy way. You do not need to manually read everything yourself. For example, if you want to look at Bolt, you do not have to go through all the files by hand. It is actually quite hard to read all of it. What you can do is convert the whole repo or part of it into LLM readable text. This means the LLM will understand it better because the information is structured properly. I would not recommend doing this with the entire repo because it is quite large and has a lot of tokens. Instead, if you want to give just the cursor folder to an LLM to question and understand the agent prompt, you can replace the hub in the GitHub link. It is a pretty amazing trick. It converts whatever you give it into LLM readable text. You can then copy it and you will also get the directory structure. Although you may not need it here, you can paste that text into any LLM you want. In this case, the estimated token count is around 7,000, which is not a lot. Even the free version of Chat GPT should be able to answer questions by looking at it. Before ending the video, I want to talk about some interesting things I found about Replet from reading their leaked system prompts. Replet is explicitly forbidden from using Docker or any kind of external virtual environment. It uses its own environment and creates a ripple which allows it to have more control over the development environment. The agent has 17 specialized tools, one of which is the VNC window application feedback tool. This means it can test GUI apps inside a VNC window. There are other tools as well like spinning up a Postgress database. Replet also has autoconfigured workflows. For example, if any application needs an API key, it automatically asks you. It prevents errors from progressing further. I noticed that Replet has gotten really good at building front ends, especially after its integration with Shad Cn UI. A big reason it excels at this is the built-in feedback loop. After it iterates something, it must verify that there are no errors. This might be why it builds front ends with far fewer errors compared to other AI coding agents. There are built-in restrictions with databases, too, which help protect your data through a safety net. That's it for this video. If you liked it, please consider subscribing. If you want us to keep making these videos and since my wallet is running a bit empty, please consider donating using the link below. Thanks as always for watching and I'll see you in the next