The sense of stealing is extremely strong, and the underlying prompt words of many commonly used AI tools such as Cursor and Manus have been exposed.

Everyone is already very familiar with AI programming tools, headed by Cursor.

I don’t know if you have a sense of crisis, but one person can write code with AI instead of three people, especially for code novices, who can make a simple application with almost no experience.

What you may not have expected is that behind these AI tools is actually the prompt word engineering. By optimizing the prompt to the extreme and maximizing the user experience, we have the Cursor.

Cursor=prompt word+engineering optimization.

Engineering optimization is to maximize the various functional operations in VS Code so that everyone can use it comfortably.

Recently, a great guy may have been bored and made public the system prompts of many AI tools including Cursor, Manus, Same.dev, Lovable, Devin, Replit Agent, Windsurf Agent and VSCode Agent .

And it was posted directly to GitHub.

The most interesting one is Devin, which adds the prompt words directly, without letting the AI ​​tell the user the prompt words.

In fact, if you want to take it off, it is complicated and simple at the same time.

There may be many ways, here is a simple one that I can think of.

You can use your own API in Cursor and see the prompt words directly on your forwarding server.

The link to the prompt words is at the end of the article. Let’s first analyze the prompt words of several commonly used ones.

Cursor Prompt

The prompt words are still very long, in order to constrain the AI, reduce hallucinations, and ensure that the content output each time is very neat and high-quality.

If you simply organize the contents of the constraints, you will find that the answers and code modifications given by the Cursor are usually in strict compliance with these "terms".

Overall positioning

You are a powerful AI coding assistant powered by Claude 3.7 Sonnet or GPT-4o, pair programming with users in the Cursor IDE to solve users' coding tasks.

Tool call rules

• Strictly follow the calling format: You must strictly follow the specified tool calling mode and ensure that all necessary parameters are provided.

  
  

• Only available tools are used: tools not explicitly provided are not called.

• Avoid mentioning tool names: When communicating with users, don’t mention the tool names, just state the actions to be performed.

  
  

• Call tools on demand: Call tools only when necessary. If the user's task is broad or the answer already exists, respond directly without calling the tool. If additional information is needed, obtain it through tool calling first.

  
  

• Explain the reason for the call in advance: Before calling the tool each time, explain the reason for the call to the user.

Code modification rules

1. Use editing tools directly:

Unless the user requests it, the code is not output directly to the user. Instead, changes are implemented using a code editing tool, which is used at most once per round.

2. Make sure the code can be run immediately:

• Combine editing operations on the same file to reduce the number of tool calls.

• When creating a codebase from scratch, create a proper dependency management file (such as requirements.txt) and add package version information and a useful README.

• When building a web app from scratch, design a beautiful, modern user interface that adheres to best UX practices.

• Do not generate long hashes or non-text codes (such as binary codes).

• Except for simple append editing or creating a new file, read the content to be edited before editing.

• Fix introduced (code inspection) errors. If you are not sure how to fix them, do not make unfounded guesses. Code inspection errors in the same file should be fixed no more than 3 times. The user will be asked for the next step on the third attempt.

• If a reasonable code edit suggestion was not applied, try reapplying the edit.

Manus

You wouldn’t have thought that Manus is actually a project with a lot of prompts, more than 200 lines of prompts.

• Core functions: It can process information, create content, solve problems, and complete tasks such as information query, code writing, document writing, and troubleshooting.

  
  

• Available tools: Covers browser operations, file system management, Shell command execution, communication interaction, and application deployment.

  
  

• Technical support: Supports multiple programming languages ​​such as JavaScript and Python, as well as common frameworks and libraries such as React and Django.

  
  

• Task processing: first analyze the needs, break down complex tasks, then formulate a plan for execution and conduct quality verification, while maintaining communication with users during the process.

  
  

• Usage restrictions: cannot access proprietary information such as internal architecture, cannot perform harmful, illegal or unauthorized operations, and have context window restrictions.

  
  

• Effective prompts: When users make requests, they should be as specific and clear as possible, provide background, and specify the format. If necessary, prompts can be optimized through iteration.

Summarize

There are many other tools, which are actually prompt word projects. Let me give you a common example. AI papers are followed by a large number of prompt words that push AI to write the paper step by step. Of course, the process is still a bit complicated, and I will have the opportunity to replicate it for you later.

The prompt word engineering is really very important. It is no exaggeration to say that it is the "operator of AI tools". For the current AI models, prompt words are still needed to constrain and guide them.

These tools add prompt words that prevent users from being informed of the prompt words. In fact, this is a cover-up. Those who need the prompt words can use them, while those who don’t need them will just use them for fun.

But it is not completely without threshold. A good tool cannot rely solely on prompts. Otherwise, if we know other people’s prompts, why are there still some AI programming software that are not easy to use? This is because the engineering optimization is not done well.